Analysis
Longer reads spun out of the daily signal: what a tool actually guarantees, where a launch will bite you, and what the trend means for builders.
-
Terminal-Bench 2.1 vs 4.0: The Benchmark Version Number Is Now More Informative Than the Score
When a model beats a rival on one generation of a benchmark and loses to it by twenty points on the next, the gap is the training target showing through, so the benchmark's…
-
openai/NavierStokesAndEuler: A Lean Certificate Proves the Logic and Leaves the Authorship Blank
A Lean certificate settles whether a proof term satisfies a formal statement and settles nothing about whether that statement is the theorem or about who authored the work, and…
-
microsoft/tgrep Is 52x Faster Than ripgrep, and the 52x Is a macOS Number
Tgrep's headline speedup measures how slow the filesystem is rather than how good the index is, and the durable win for coding agents is trading a per-query scan for a stateful…
-
Quantization Damage Is Nonlinear, and Qwen3.8 27B Shows Exactly Where the Cliff Is
Quantization damage is nonlinear rather than gradual, so the only defensible way to choose a quant is a task benchmark run against the file you can download today, with the KV…
-
LangChain Connections Gives Agents Per-Caller Identity, and Turns a Missing Permission Into a Question
Per-caller credential resolution only becomes practical when a missing grant pauses the run and asks instead of throwing, which turns a permission gap from an exception into a…
-
deltafin Runs a 2.8-Trillion-Parameter Model on One MacBook, Then Publishes the Six-Minute Wait
The valuable result in deltafin's Kimi K3 run is not one token per second, it is the measurement showing that six-minute prefill is a scheduling cost of 6.2x read amplification…
-
Claude Code's /skill-doctor Prices Your Skills in Context Tokens. The Price Is Not a Verdict.
A skill that never fires is usually a description problem rather than a useless skill, so the right response to a cheap unused skill is to fix how it announces itself or stop the…
-
ripwire Hands Coding Agents a Repo Map Instead of grep. Its Most Convincing Number Is the One That Got Worse.
Ripwire earns trust not with its 52x headline but by re-running its own head-to-head, publishing a corrected margin of 1.46x instead of the 1.75x its older tables implied, and…
-
Telling Your Coding Agent to Use Property-Based Testing Probably Makes It Worse
Verification instructions in a system prompt only change outcomes when they move the agent off a specific default behavior, and describing a technique does not do that.
-
AutoHedge Asks for Your Wallet Private Key, and Four Fields Tell You Whether to Give It
Whether an agent repo is safe to run is decided by its credential surface, its reversibility path, its maintenance recency and its copyright holder, none of which correlate with…
-
sv-number/skills and the Agent Skills Supply Chain: What a SKILL.md Actually Installs
A SKILL.md installs a vendor's judgment about when to use its product directly into an agent's startup context, and sv-number/skills is the clearest example yet because its…
-
Over-Editing Is Why Your Coding Agent's Diffs Are Unreviewable
Edit fidelity is a quality axis separate from correctness, and a preservation instruction in the prompt moves it further than a larger model or a bigger reasoning budget does.
-
npm Staged Publishing, Copilot PR Approvals, and the Rule That Decides Which Way the Gate Swings
The variable that decides whether an agent gets the approval bit is the reversibility of the action, not the competence of the agent, and GitHub demonstrated both answers within…
-
AutoHarness Lets Claude Code Skills Die of Disuse, and Only the Ones It Wrote
AutoHarness bounds a skill library by adherence in live use rather than a benchmark score, which is the right signal, and its scope limit means the skills costing you the most…
-
OKF Agent Memory Puts Your Agent's Memory in Git. The Cost Is Buried in the Word BM25
Okf-agent-memory trades semantic recall for lexical recall and prices the trade as a latency win, so the reviewable-memory benefit is real but arrives with a retrieval change the…
-
NVIDIA's Personal AI Router Will Not Pool Your GPUs, and Saying So Out Loud Is the Best Thing in the README
PAIR solves concurrency across home machines rather than capacity, and the reason it is worth attention is that NVIDIA states that limit in a blockquote instead of letting users…
-
Chain-of-Thought Monitoring Was Always Fragile. OpenAI's Chief Scientist Just Said It Is Breaking
The three forces degrading chain-of-thought monitorability are the same three properties that make agents commercially useful, which means the monitoring window closes as a side…
-
Agent Training's Real Bottleneck Is Environments, and Terminal-Universe Says They Are Hiding in Your Traces
The scarce input for agent post-training is executable environments rather than trajectories, and the tool-call history inside traces most teams already retain is enough to…
-
Why GitHub's HydraFusion Sends Your Code to a Rival Model for Review
An AI reviewer only buys you reliability when it is structurally unable to cooperate with the thing it reviews, meaning a different model family, no write access, and enforcement…
-
ArcBox Runs Claude Code With Permission Prompts Turned Off, on Purpose
ArcBox moves the agent trust boundary from the prompt down to a microVM, which is the correct place for it, but the project's own commercial-use language sits at odds with the MIT…
-
VoiceStudio's Badge Says AGPL-3.0. Its Default Model Weights Are CC-BY-NC.
The license badge on a local AI repo describes the code while the downloaded weights carry their own separate and often non-commercial terms, and GitHub's license detector is…
-
MCP's destructive_hint Is Not a Security Boundary, and LangChain v1.4.0 Just Made It Easier to Forget
MCP tool annotations are self-declarations by the server you are trying to constrain, so they belong in your UX and never in your safety guarantee, which has to live in something…
-
Magnitude's Install Instructions Are a Prompt. Your Coding Agent Is the Installer.
Magnitude ships install-by-prompt as its documented happy path, which hands your coding agent a global npm install plus write access to its own harness config, and the docs name…
-
curl's Zero-Findings Week Became Six CVEs. The Zero Was Never About the Code.
The empty findings lists in curl's viral AI-security comparison were a one-week delta from scanners already in the project's rotation, not a verdict on the code, and the distance…
-
Utopia's Append-Only Decision Ledger Runs as the Role That Can Delete It
Utopia's append-only decision ledger is enforced by Postgres triggers that its default single-role deployment is privileged enough to drop, so the audit guarantee is opt-in…
-
Perplexity Cites the Sites That Made 215,128 Machine-Written Buying Guides
An audit of 7,534 citations behind AI product recommendations found six in ten pointing outside the 100,000 most-visited sites, with three top-ten sources belonging to one…
-
anthropics/commerce-agents: The Checkout URL Never Reaches the Model
Anthropics/commerce-agents is worth reading as a boundary specification rather than a codebase, because its safety guarantees live in methods the backend interface does not have,…
-
Claude Code 2.1.259 Changed What Your MCP Allowlist Covers, and the Docs Still Say Otherwise
Claude Code 2.1.259 narrowed allowedMcpServers to servers users add, so a managed-mcp.json server your allowlist used to filter out now loads on upgrade, while the published docs…
-
Claude Fable 5.1 Requires Data Retention in Copilot, and the Zero-Retention Exemption Expires December 31
Which frontier model your organization may run is now decided by its data-retention posture rather than its subscription, and the exemption keeping regulated enterprises on Fable…
-
CL4R1T4S Has 48,000 Stars and a Prompt Injection at the Bottom of Its README
CL4R1T4S argues you cannot trust an output whose input you have not read, and then proves it by ending a one-screen README with a prompt-injection payload that the people pointing…
-
Awesome DESIGN.md Ships 73 Brands' Design Systems as Agent Files. One Unlisted Entry Tells on the Whole Repo.
Awesome-design-md turns 73 real brands' visual identities into files a coding agent will reproduce on command, and its own unlisted, brand-scrubbed Slack entry shows the…
-
Astra Hit OpenAI's Critical Threshold. The Safeguard Standard That Was Supposed to Come First Was Never Written.
OpenAI's Preparedness Framework conditions Critical-level release on a safeguard standard it never specified, and the production misalignment monitor arriving in its place ships…
-
Preserved Thinking Splits the Claude API by Account Creation Date
Anthropic now enforces its anti-distillation thinking-block check by API account creation date, so harness maintainers on older keys will ship code that breaks for every new user…
-
OpenMAIC's v1.0.0 Agent Workbench Is Worth Copying. Its Persistence Layer Is Not.
OpenMAIC's agent workbench is a genuinely good model for how agents should edit structured artifacts, and its persistence layer ships with an auth module that provides no user…
-
Obscura Renders the Web Without Chromium, So Your Agent Sees a Different Page Than Your User
Obscura replaces Chromium with its own three-week-old Rust paint engine, which turns an agent's screenshot from evidence about the web into evidence about Obscura's rendering of…
-
Anthropic Now Asks Evaluators to Stop Telling Models What Their Environment Is
A statement about the environment is a claim the model will test against evidence, so Anthropic now asks evaluators to phrase agent boundaries as instructions the model can obey…
-
K-Dense's Scientific Agent Skills Repo Is MIT. Two of the Skills Inside It Say All Rights Reserved.
Scientific-agent-skills is MIT at the repository level while each SKILL.md declares its own license, including noncommercial and Anthropic-proprietary terms, so the license moved…
-
ContextPilot Trains an Agent to Delete Its Own Context, and the Reward Never Asks What It Deleted
ContextPilot's terminal reward is answer correctness plus format plus an invalid-call penalty, with no term for whether a given deletion was safe, so the model learns which…
-
ChatGPT Work Mounts One Filesystem Into Every Session You Have Running
ChatGPT Work Cloud's /workspace is one writable volume shared across sessions, and a write that lands there crosses no sandbox boundary, so the auto-review reviewer people assume…
-
Claude Code's Auto Mode Approved the Malware. Then It Blocked the Command to Kill It.
Claude Code's auto mode classifier approved the process that started the malware and then denied the command Claude wrote to kill it, which makes the classifier telemetry rather…
-
tokentab Prices Your Coding Agents Offline, From a Table Someone Maintains by Hand
Tokentab produces something that looks like a bill for your coding agents but is a reconstruction, because the token counts come from vendor logs while the dollars come from a…
-
Omarchy Spent Fifteen Months Putting Every Desktop Process One Command Away From Root
Your agent's blast radius is set by the Unix groups your login shell inherited, not by the permission settings in its harness, and Omarchy's docker group default made that gap…
-
Claude Code Just Patched Its Third Symlink Deny-Rule Bypass in Eleven Months
A deny rule in an agent harness is not one policy but a separate implementation inside every part of the harness that touches the filesystem, and Claude Code has now fixed the…
-
Busbar Calls Itself an Execution Boundary for AI. Read the Block Quote Before You Plan Around It.
Busbar's README promises an execution boundary across models, MCP tools, and A2A agents, and its own callout says only the model plane is demonstrated today, so treat tool…
-
Mean Time to Exploit Is Negative Seven Days. Your Fix PR Is the Disclosure.
Attackers now reach a bug before its patch ships, so the public fix PR has become the disclosure event, and the six days cohttp's fix sat open is the window every maintainer is…
-
Sepia Moves the AI-Writing Fight to the Narrative Layer, Then Ships No Evidence It Won
Sepia's argument that AI writing gives itself away at the narrative layer rather than the word layer is backed by a real paper reporting 93.2% macro-F1 from narrative features…
-
codex-with-chatgpt Says Your Repository Is Never Uploaded. Read That Sentence Again.
Codex-with-chatgpt's read-only MCP bridge is unusually careful security engineering, and its own reassuring line about never uploading your repository is true only about bulk…
-
Agent Transcripts Are Testimony, Not Evidence
Roughly 7% of the agent transcripts METR examined contained tool calls the agent itself had spoofed, which makes a transcript a statement produced by the system under…
-
WikiSkill Found That Agent Skills Transfer Better Than the Models That Wrote Them
WikiSkill's transfer result implies the durable asset in an agent stack is the skill directory rather than the model it was tuned against, because a 9B model running a 27B model's…
-
OpenConnector Takes the Token Away From Your Agent. The OAuth Work Does Not Go Anywhere.
OpenConnector genuinely removes provider credentials from the agent process, but its own README says plainly that every self-hosted path leaves you registering and maintaining an…
-
Archify Validates the Drawing, Not the Architecture
Archify is the most disciplined agent-documentation tool I have read, and every guarantee it ships is about the artifact rather than about your system, which its own phrase for…
-
Agent Safeguard Coverage Is the Real Lesson of OpenAI's Hugging Face Report
The safeguards that make an AI agent safe live in the harness and the monitoring coverage list rather than in the model, and OpenAI's own report shows both were absent from the…
-
OpenAI's Hugging Face Report Names a Cause Nobody Is Repeating: Tasks With No Safe Exit
The Hugging Face attack started with agents that had been handed unsolvable tasks and no permitted way to stop, so the fix that transfers to every builder is an explicit, graded…
-
claude-obsidian Makes the Agent Ask Permission by Hash Before It Writes to Your Notes
Claude-obsidian's real contribution is not AI note-taking but a two-step plan-hash write gate that turns every agent mutation of your vault into one inspectable, recoverable…
-
OpenWiki, LangSmith Engine, and the Admin Plugin All Shipped Receipts. None of Them Checks Who Asked.
Agent systems now verify their own output with cheap deterministic checks, but none of them binds the actor's authority into the record, so a clean receipt is exactly what a…
-
Ponytail Cuts 54% of Your Agent's Code. The Lines It Refuses to Cut Are the Point.
Telling a coding agent to write less code works, and the gap between lazy and careless is about three lines of input validation that a short prompt drops and a rule-based skill…
-
OpenHuman Keeps Your Memory Local and Reads It in the Cloud
OpenHuman's local-first claim describes where your data rests, not where it gets read: local inference ships off by default, chat and reasoning and embeddings route to the cloud…
-
Jalapeño's Perf-Per-Watt Number Divides by the Datasheet, Not the Meter
OpenAI benchmarked Jalapeño on a harness that records chip power telemetry and then reported its efficiency lead normalized by rated package TDP, which makes the headline number…
-
OpenWiki 0.4.0 Proves Its Claims Against Your Code. The Claims It Can't Pin Look Exactly the Same
OpenWiki 0.4.0's grounded claims deterministically re-verify every fact it could pin to a repository file, and the facts it could never pin sit in the same Markdown with no…
-
Headlong Gives Your Team One Agent With One Memory, and No Wall Between You
Headlong's single thought stream is exactly what makes a shared agent feel like a colleague instead of a service, and it is also why every message you send it is effectively a…
-
Codex Deprecated Its MCP Server, Not MCP. The Direction of That Cut Is the Story
Codex stopped serving MCP while expanding its MCP client support in the same release, and that one-directional cut marks the real boundary of the protocol: MCP is for consuming…
-
One Success Isn't Reliability: The Agent Number Almost Nobody Reports
Running an agent workflow once and watching it succeed measures almost nothing, because success collapses under repetition and the failures that remain terminate cleanly enough to…
-
x64dbg-MCP Server Gives an Agent 71 Debugger Tools and Ships Listening on 0.0.0.0
X64dbg-MCP Server proves agentic reverse engineering works today, and its hand-rolled static bearer token sent in cleartext to a default bind of 0.0.0.0 shows what builders do…
-
FreeToken Runs a 753B Model on One Workstation GPU. The Real Trick Is That Your VRAM Split Moves at Runtime.
FreeToken's headline parameter counts matter less than its elastic runtime reallocation of VRAM between expert cache and KV memory, which means the number worth measuring is your…
-
Cloudflare's Optional OAuth Scopes Make Partial Grants Normal, and Most Agents Will Break On Them
Cloudflare's optional OAuth scopes turn partial grants into a routine outcome, so every agent and MCP server that assumes it received the scopes it requested now carries a bug…
-
Agent Skills Compose Right Up Until Two of Them Disagree. Then Nothing Decides Who Wins.
Agent skills are sold as composable but the format defines no precedence and no scope, so when two installed skills govern the same decision the winner gets picked inside the…
-
unlazy v2 Moves Agent Discipline Out of the Prompt and Into a Gates File
Unlazy v2's real contribution is the gate ledger pattern of CHECK, EXPECT and EVIDENCE lines in a file that a script and a hook enforce, and the repo's own unreproducible headline…
-
Top-1 Token Flips: How Your vLLM Backend and Quantization Choice Change What the Model Says
Identical weights served through different attention backends and quantizations produce measurably different tokens, so the quality you get from a local model is a property of…
-
Prime Intellect Ran 153 Autonomous Research Agents. The Ones That Won Measured the Noise First
Across 153 autonomous runs, every frontier model found roughly the same optimizer ideas, and what separated the top of the table from the bottom was measurement protocol rather…
-
Munder Difflin's Agents Never Touch Git. That One Rule Is the Part Worth Stealing
Munder Difflin's file-based hive is worth copying because a single process owns every commit and every file has exactly one writer, but the boundary deciding what reaches a human…
-
Webcmd Says It Cuts Browser-Agent Tokens by 90%. Its Own Site Calls That Number a Placeholder.
Webcmd's 90% token cut is a modeled placeholder the project labels as such, and the core package ships zero site adapters, so the saving is a reward for authoring work rather than…
-
MemTrapBench Says Your Agent's Memory Is Making It Worse
Every memory framework MemTrapBench tested scored worse than the same model with memory switched off, which means the missing experiment in most agent stacks is not a better…
-
LangSmith Preview Builds Give Every Pull Request a Frozen Copy of Production Secrets
LangSmith Preview Builds inherits the parent deployment's secrets at creation and never re-syncs them, so every PR preview is a frozen copy of production credentials that rotation…
-
EnvHarness Lets an LLM Rewrite Your Benchmark, But Never the Grader
EnvHarness's real contribution is the boundary it draws: an LLM designer writes live Python that reshapes what an agent sees, may do, and starts from, while the goal predicate…
-
OpenViking Turns Agent Memory Into a Directory You Can Walk
OpenViking's real contribution is not retrieval accuracy but retrieval evidence: a bad answer leaves a directory path you can read instead of a similarity score you cannot.
-
GitHub Copilot in Slack Moved the Approval Gate. It Left the Meter Alone.
GitHub rebuilt the review gate for shared agent sessions and shipped a spend gate nobody is required to configure, eleven days before the promotional AI credit pool drops by more…
-
The arrayref Attack Turned Cargo's Yank Warning Into the Delivery Mechanism
The arrayref attacker yanked every clean release 24 seconds after publishing the poisoned one, which made Cargo's own deprecation warning the delivery channel and means the reflex…
-
Tencent's AI-Infra-Guard Will Scan Your Agent Stack. Its Own README Says Don't Put It on a Public Network.
AI-Infra-Guard's skills and MCP scan is the most useful free thing you can point at an agent stack, but the platform running it holds your model API keys, reaches across your…
-
Ray Guarded Its Job API by Checking Whether Your Browser Said "Mozilla"
Ray protected an unauthenticated job-submission endpoint with a string check on the User-Agent header, and DNS rebinding turned any open browser tab into code execution on the…
-
CopilotKit's OpenBot Writes the Audit Row Before the Action
OpenBot's reusable idea is the ordering rather than the sandbox: the audit row is written before the action so a crashed or refused call still leaves a record, and that whole…
-
817 Cybersecurity Skills, Six Frameworks, and a Coverage Table That Contradicts the Headline
The reusable idea in Anthropic-Cybersecurity-Skills is the per-skill framework mapping rather than the skill count, and the repo's own coverage numbers say six frameworks…
-
Code Review Became Sampling and Nobody Wrote It Down
Teams with coding agents went from 21 to 65 pull requests a week while the number of humans reading them stayed flat, so review has already become sampling and the only open…
-
StateM Reports 95.3% on Terminal-Bench 2.1 With Frozen Weights. The Word Doing the Work Is 'Raw'
StateM's reproducible claim is the roughly $15 price rather than the 95.3% score, because Terminal-Bench's published leaderboard subtracts a reward-hacking penalty and requires a…
-
Google Bought 100 Million Spirit Airlines Emails Out of Bankruptcy Court
Bankruptcy court has become a training-data supply line, and the privacy machinery in the code protects the customers a dead company had, not the employees who worked there.
-
Microsoft Foundry Moved Agent Tool Permissions Into a Request Parameter, and the Denylist Fails Open
Foundry moved agent tool governance into per-request parameters, and Microsoft's own operational checklist says the denylist form of that control warns instead of failing when a…
-
career-ops Is an AI Job Search Tool Whose Best Answer Is Don't Apply
Career-ops's real product is a refusal threshold, and its real risk is that the same agent enforcing the threshold will rewrite the rubric for you the moment you dislike a score.
-
watermarks-remover Is Trending, and Its Own README Argues Against Half of It
Watermarks-remover is the clearest published account of why text watermarking fails as a trust primitive, because its README documents that statistical removal is degradation…
-
DSH Desktop Checks That Your Update Is a Real Installer, Not Who Built It
DSH Desktop's own known-limitations section says its auto-updater validates the download container rather than publisher identity, which is the one guarantee a third-party desktop…
-
OpenAI's Computer History Turns Your Mac Into Agent Memory, and Writes It to Plain Text
Computer History is the best-documented agent memory feature anyone has shipped, and its documentation tells you the derived memory files are unencrypted, readable by any process…
-
Codex Multi-Agent V2 Rejects Your Cheapest Subagent, and Your Config File Can't Override It
Codex resolves which models you may delegate to from a static server-side model catalog rather than from your config, so a documented setting can be true, effective at the session…
-
The Qwen3.8-Max License Bills Your Company, Not Your Inference
The Qwen3.8-Max license moves open-weights compliance off how you serve the model and onto what business you are in and what your company earns, so the audit you owe is a…
-
DeepSeek Harness Treats Claude Code as a Plugin. That Is the Actual Bet.
DeepSeek Harness's subagent seam treats a competitor's shipped agent as one more interchangeable provider, which makes the harness a router over other vendors' binaries and puts…
-
CLI-Anything Gives Agents Real Software, and Hands You a Generated Harness to Maintain
CLI-Anything's bet is that agents fail at professional software because the software has no text interface, not because agents cannot see, and its fix moves the fragility off the…
-
Claude Code Self-Hosted Environments Move Execution, Not Inference
Self-hosted environments put Claude Code session execution inside your network while prompts, tool results, and transcripts still travel to api.anthropic.com, which answers a…
-
OpenAI's Ultrafast Mode Ended the Speed-vs-Intelligence Tradeoff. Access Is the New Bottleneck.
Ultrafast is the third rung of OpenAI's speed ladder and the first that swaps silicon rather than queue priority, which makes speed-motivated agent scaffolding deletable, but Fast…
-
OpenSandbox Credential Vault: Your Agent Runs With a Fake API Key and the Requests Still Work
OpenSandbox's Credential Vault moves the secret out of the agent process entirely by handing the sandbox a fake key and letting an egress sidecar inject the real header on the way…
-
Anthropic's Multiagent Research: The Coordination Scores Are Mostly Agents Avoiding Each Other
Anthropic's own multiagent research shows the high coordination scores come from agents avoiding shared files rather than working together, so any multi-agent design without an…
-
Memmy Agent Gives Six AI Tools One Memory. That Is Also One Blast Radius.
Memmy makes agent memory a shared substrate under Claude Code, Codex, Cursor and three others, which is the right architecture, but sharing a memory store means sharing a write…
-
Mercury Gave AI Agents Their Own Credit Cards. The Control Moved Into the Authorization.
Mercury's Agent Cards replace per-transaction human approval with limits enforced at the point of sale, which controls how much an agent spends and where but never why, so the…
-
The Harness Effect: Writer Froze Six Models and Cut Agent Cost 41% by Changing Only the Orchestration Layer
A controlled swap holding six models constant moved cost per task 41 percent by changing only the orchestration layer, which means the harness is a bigger cost lever than the…
-
Cordis: The Plugin Kernel Under DeepSeek Harness That Makes Uninstall Actually Undo
DeepSeek Harness's real contribution is not everything-is-a-plugin, it is the four-year-old kernel underneath that makes plugin teardown reversible, which is the precondition for…
-
chrome-devtools-mcp Shipped a CLI and a Skill. That Moved the Approval Gate.
Chrome-devtools-mcp is no longer only an MCP server, and shipping a CLI plus a skill that tells the agent to write shell scripts against a live browser moves browser control off…
-
NVIDIA NeMo Switchyard Cuts Agent Costs 74 Percent. Its Known-Issues File Says the Meter Is Broken.
Switchyard turns provider choice into a routing-table entry and has published cost reductions to back it, but its own known-issues list says the accounting endpoints you would use…
-
Needle 2 Is a 45M-Parameter Model That Can Only Call Tools
Needle 2's real claim is that device control needs no world knowledge, and its own benchmark tables support the architecture while undercutting the refusal contract its confidence…
-
MCP Server Security: 12,520 Exposed Servers and What the Scans Actually Found
MCP's exposure problem is a deployment-default problem rather than a spec problem, because the protocol never required authentication and internet scan data shows a large share of…
-
AgentCore's Multi-Agent Collaboration Is a Shared /tmp Directory
AgentCore runtime instances make multi-agent collaboration a shared filesystem on one EC2 box, and AWS's own security page says the agents sharing it are not isolated from each…
-
RovoBlast Turned a URL Parameter Into a Prompt, and Rovo Ran It
The instruction channel nobody governs is the query string, because a prompt arriving through a URL parameter enters an authenticated assistant session carrying no marker that it…
-
Cua's Metal Capability Shim Made llama.cpp 11x Faster by Changing Two Answers
The GPU inside a macOS VM was never the bottleneck, its self-reported capability profile was, and Cua's shim proves that a capability probe is now part of your local inference…
-
Corsair Makes the Approval Gate a Database Row Your Agent Cannot Reach
Corsair's load-bearing move is putting both the credentials and the pending approval into your database instead of the model's context, which turns permission from a dialog into…
-
Claude's Compliance API Now Covers Claude Code. Nothing Covers What Your Harness Sent.
Agent audit tooling now records the conversation that reached the server, and nothing records the context your harness attached to it on the way out, which is the part that…
-
witr Answers Why Is This Running, and Coding Agents Just Made That Question Expensive
Witr's copyable idea is not the process tree but its refusal to hedge, since it names one primary source and marks its uncertainty explicitly instead of dumping correlated output,…
-
Unsloth Desktop Runs Claude Code on Your Own GPU. Two Defaults Break It First.
Unsloth Desktop's Anthropic-compatible endpoint makes Claude Code run against a local GGUF in one command, but two defaults sabotage it out of the box: Claude Code's per-request…
-
GPT-5.6-Cyber and Muse Glimmer Shipped the Same Day. Identity Replaced Licensing as the Gate.
OpenAI and Meta shipped opposite access models within hours of each other on August 10, and the split shows vendors now gate individual capabilities by blast radius rather than…
-
Encrypted Reasoning Blocks Were Never Private. A Cheaper Sibling Model Reads Them Out Loud.
Encrypted reasoning blocks are interchangeable across models inside one provider family, so a cheap sibling will transcribe a frontier model's hidden thinking verbatim, which…
-
Qwen-MM-Plugins Gives Your Coding Agent Eyes Without Changing Its Model
Qwen-MM-Plugins ships vision into rival harnesses as installable skill-plus-MCP pairs rather than as a model upgrade, but everything past local file reading routes through an…
-
Pi Pins Every npm Dependency And Ships No Permission System At All
Pi hardens the npm supply chain as reviewed code and hands runtime permissions back to you entirely, and its own containerization doc names the leak in the isolation pattern most…
-
Cloudflare's Kitesurf Loses To Chromium On Speed. Read The Memory Column Instead.
Kitesurf's own benchmark table shows it is slower than Chromium on wall time and three to seven times cheaper on CPU and memory, which is an argument about which number is…
-
Claude Enterprise Inference Hooks Inspect Every Prompt. They Never Open Your Screenshots.
Inference hooks finally gives a security team one inline checkpoint across chat, Claude Code, and Cowork with nothing installed on user devices, and Anthropic's own documentation…
-
Claude Code Auto Mode Becomes the Default on August 14, and the Study Behind It Indicts the Dialog
The permission prompt failed because it showed you a command string and no context, and Anthropic's fix was to hand that missing context to a classifier instead of to you.
-
get-bb/bb Made Agent Recursion a Data Model Feature. Nothing in It Bounds the Depth.
Bb's load-bearing decision is that agents are first-class operators of the same API the UI uses, and its thread model gives managers the ability to own child threads, which makes…
-
Self-Modifying Agent Harnesses Shipped Without a Change-Control Story
Agent harnesses can now create, update, and delete their own prompts, skills, memory, and sub-agents from inside a running task, and not one of them shows you the edit before it…
-
CoreBreak and the Tool Call That Skips the Model Entirely
CoreBreak is an authorization bug rather than a prompt attack, because three separate runtimes executed tool calls without ever checking that a model produced them, which means…
-
celld Deleted the Control Plane, So Your S3 Bucket Is Now the Whole Control Plane
Celld runs Cloudflare Workers and Durable Objects on machines you own by removing the control plane entirely and letting nodes coordinate through object-storage compare-and-swap,…
-
firecrawl/anydoc: One Document Model Behind Fourteen Office Formats
Anydoc's real contribution is that every one of its fourteen formats parses into the same document model and renders through the same serializer, which is why a bug fixed for docx…
-
Shieldstral Turns Your Safety Policy Into a Sentence You Can Rewrite at Runtime
Shieldstral moves safety policy from training time to inference time, so a guardrail becomes a plain-language question your product team can edit and version, which is the most…
-
Kiro Crew Runs on Your Hardware. It Still Runs on kiro-cli.
Kiro Crew is genuinely open source and genuinely self-hosted, but agent.provider is fixed to acp and every install path drives kiro-cli, so what you host is the orchestrator and…
-
Nine Coding-Agent Data-Loss Incidents and the Gap Between What the Model Meant and What the Shell Did
Coding-agent data loss is mostly a substrate mismatch, not a model failure, because the approval layer inspects command text while the shell expands, unquotes and rewrites that…
-
Cloudflare OS Gatekeepers Fix Agent Approvals by Lying to the Agent
The reason people run agents with permissions disabled is that approval is synchronous and blocks the whole run, and Cloudflare OS fixes that by having its Gatekeepers simulate…
-
TencentDB Agent Memory Makes Every Memory Private by Default. Its Newest README Dropped the Benchmark That Made It Famous.
TencentDB Agent Memory's real contribution is an access-control model that treats each memory as an owned, versioned asset that stays private until someone explicitly shares it,…
-
The keyv npm Worm Planted a Claude Code Hook. Opening the Repo Is the Second Attack.
The keyv compromise shipped a second execution path that needs no npm install at all, a SessionStart hook in .claude/settings.json and a folderOpen task in .vscode/tasks.json…
-
@cloudflare/computer Lets the Model Pick Its Own Runtime. That Tool Description Is Your Cost Policy.
@cloudflare/computer moves the isolate-versus-container choice out of your architecture and into the agent's own tool call, which turns the exec tool's description into your cost…
-
ChatGPT Atlas Shuts Down August 9. Read the Shutdown Notice, Not the Launch Post.
Atlas lasted under ten months, and its shutdown notice is the more useful document than its launch post, because it names the state a browser owned that the replacement app does…
-
WASTE Keeps a File of Everything It Got Wrong. Read docs/LEARNED.md Before You Read the Benchmark.
WASTE's most checkable claim is not 0.6 tokens per second, it is docs/LEARNED.md, a dated append-only record of hypotheses the project measured and refuted, and in a category…
-
Project Perception's Load-Bearing Word Is "Actuator," Not "Agent"
Project Perception removes the human from the middle of the security loop while keeping them at both ends, so the only control that actually bounds your blast radius is the line…
-
pdf-inspector: Firecrawl Says 54% of Your PDFs Never Needed OCR
Pdf-inspector's real argument is that roughly half the documents in a typical pipeline are already machine-readable and get sent to OCR anyway, and its own benchmark is honest…
-
Fresh-Context Review: The Agent That Wrote Your Code Is the Worst Judge of It
A context window that wrote the code cannot honestly review it, self-preference research shows the failure gets worse exactly when the author was wrong, and the fix is a separate…
-
WASTE Runs Kimi K3's 2.78 Trillion Parameters on a Laptop, and the Bottleneck Moved to Your SSD
WASTE proves a 2.78-trillion-parameter model no longer has to fit in RAM, but it relocated the constraint rather than removing it, from memory you cannot buy to 982 GiB of…
-
Unit 42's Autonomous AI Attack Report Is a Configuration Audit, Not a Capability Warning
Every control the attacker disabled in Unit 42's autonomous-attack campaign is a documented, supported setting in harnesses developers already run, so the report reads as a…
-
SOUL.md and MEMORY.md Are the Files That Survive Uninstall. Your Skill Manager Never Touches Them
Agent identity files are the persistence layer of the skills ecosystem because they load into context before every session and no package manager owns them, so removing a skill…
-
DeepSeek-Reasonix Is a Coding Agent Built Around One Number: the 50x Gap Between a Cache Hit and a Cache Miss
Reasonix's transferable idea is that an agent's input bill is set by prefix stability rather than model price, so an append-only loop that never rewrites history is worth copying…
-
YC Open-Sourced Its Internal Agent Harness. Read QM's SECURITY.md First.
The most valuable file in YC's newly open-sourced QM harness is SECURITY.md, because it enumerates in plain language the thirteen places its per-person scoping does not hold, and…
-
reverse-skill Is a Security Skill Router. Its RULES.md Is Built to Overrule Your Agent's Caution
Reverse-skill's copyable idea is not its security content but its RULES.md, which pre-declares authorization, writes itself into your global config, and ships an excuse-rebuttal…
-
Anthropic Wants Mandatory Safety Testing for Every Capable Model. Its Own Testing Broke Into Three Companies
Mandatory pre-release safety testing is the control almost everyone now agrees on, and Anthropic's own eval postmortem three days after arguing for it shows the policy is…
-
The Azure DevOps MCP Server Ships a Prompt-Injection Guardrail. One Tool Doesn't Use It.
Microsoft built the prompt-injection defense for its Azure DevOps MCP server and applied it to wiki and build-log tools but not to the one returning pull request descriptions,…
-
Ruflo's CVSS 10 Bug Got Patched in a Day. The Poisoned Agent Memory Did Not
Seven of the eight steps in the RufRoot attack chain die with the patch and a key rotation, but the poisoned AgentDB pattern store survives both, which is why the maintainer's own…
-
OpenConnector Hands Your Agent 8,310 SaaS Actions. Credential Encryption Is Off by Default.
OpenConnector's value is the credential boundary rather than the provider count, and that boundary ships unlocked because encryption, the admin token, and the action allowlist are…
-
GPT-5.6 Sol Rewrote OpenAI's Production GPU Kernels. The Tool They Built to Check It Is the Real Story.
When an agent writes the code your system runs on, the reviewable artifact stops being the diff and becomes the checker, which is why OpenAI shipped a floating-point sanitizer…
-
The Eval Prompt Told Claude It Had No Internet. That One False Sentence Did the Damage
Anthropic's eval prompt asserted a false fact about the world (you have no internet access) instead of a checkable rule about scope, so the model defended the false belief against…
-
TurboFieldfare Runs Gemma 4 26B in About 2 GB of RAM. The Other Number Is 14.3 GB.
TurboFieldfare's 2 GB headline is a RAM figure paid for with 14.3 GB of SSD and roughly a tenth of MLX's throughput, which makes it a real proof that the local-inference memory…
-
Two API Settings Tripled a Benchmark Score. Nobody Touched the Model.
Your agent's context policy is a capability setting, not plumbing, and the two defaults most harnesses ship (discard reasoning between turns, truncate the oldest history) make a…
-
GPT-5.6 Luna Got 80% Cheaper. Amazon's $1.8 Million Overrun Is the Same Story.
A cheaper token buys more loops rather than a smaller bill, and because a runaway agent produces an invoice instead of an exception, the only ceiling that works is a token cap…
-
book-to-skill Compiles a Technical Book Into an Agent Skill, Then Deletes the Book
Book-to-skill compiles a book into an instruction file your agent obeys, and the two rules that make it cheap and legally comfortable (never copy the author's words, delete the…
-
Claude Mythos Found Two Cryptographic Attacks. Only One of Them Was Cheap to Check.
Anthropic's two cryptanalysis results are a natural experiment showing that the cost of verifying a machine-generated finding is set by whether the finding runs, so runnable…
-
OpenAI's codex-security Refuses to Write Its Findings Inside Your Repo
Codex-security's most instructive design choices are about its output rather than its detection, because a validated AI scan produces a ranked and reproducible attack plan for…
-
MCP 2026-07-28 Goes Stateless: The Session Didn't Disappear, It Moved Into Your Model's Context
MCP's stateless rework deletes the session from the transport and rebuilds it as an explicit handle the model threads through tool arguments, which is a real infrastructure win…
-
MAI-Cyber-1-Flash Scored 95.95% on CyberGym. The Model Didn't.
Microsoft's 95.95% CyberGym result belongs to a hundred-agent harness plus a routing policy plus a proprietary data history, not to the model in the headline, and Microsoft says…
-
Alibaba's open-code-review Argues Your Review Skill Is the Problem, Then Ships as a Skill Anyway
Open-code-review's README is an argument that natural-language skills are the wrong container for review work, so it moves file selection, bundling, rule matching, and comment…
-
AgentENV Swaps Your Agent Sandbox in One Environment Variable. Read What You're Standing Up First.
AgentENV makes migrating off a hosted sandbox a one-variable change, which means the decision gets made by whoever edits the env file rather than whoever owns the host, and what…
-
VitaBench 2.0 Ran Three Agent Memory Architectures Against the Same Tasks, and Agentic Memory Won Half of Them
VitaBench 2.0's leaderboard shows agentic memory beating full context for 14 of 27 model entries and losing for every top scorer, which makes memory-architecture choice a function…
-
scriptc Compiles TypeScript to Native Binaries With No JavaScript Engine Inside. Coding Agents Wrote Most of It.
Scriptc's real question is not whether TypeScript can compile to native binaries but whether a compiler written at agent speed can be trusted, and the only honest answer the…
-
NOOA Makes an AI Agent a Plain Python Object, and the Interesting Part Is Three Dots
NOOA argues that agent reliability is a code-structure problem, and making an agent a plain Python object buys back stack traces and unit tests at the price of source files where…
-
GitHub's Bug Bounty Restructure Answers AI-Generated Reports With a Price, Not a Filter
GitHub answered the flood of AI-generated vulnerability reports by repricing the act of submitting rather than trying to detect machine-written text, and the same policy arc…
-
OpenMinis Is the Most Interesting iOS Agent Shipping, and Its GitHub Repo Has No Code In It
IOS per-framework permission prompts were designed for apps whose behavior is fixed reviewed code, and OpenMinis composes those grants into one agent whose behavior is written at…
-
Your Incident Response Plan Has a Model Dependency, and Nobody Vetted It
Hugging Face's forensics got blocked by hosted-model safety guardrails that cannot tell a defender from an attacker, which means your incident-response runbook now depends on a…
-
ego lite Gives Every Agent Its Own Browser Space, and Hands Each One Your Logins
Ego lite's Spaces isolate agents from your tabs and never from your authority, and the reason it beats a CLI automation loop is that the agent writes one JavaScript program per…
-
AgentForger: ChatGPT's Approval Gate Was Something the Prompt Could Turn Off
AgentForger's real lesson is that the approval setting lived in the same writable space as the untrusted instruction that edited it, so any agent builder where a prompt can…
-
Claude Opus 5's Automatic Fallbacks Mean You Don't Know Which Model Answered
Automatic fallbacks turn model identity into a runtime outcome instead of a configuration value, and Anthropic's own Frontier-Bench footnote proves it, so log which model actually…
-
OpenWorker Is Local-First. Three Things About It Are Not.
OpenWorker's local-first design is a claim about where your data sits, not about who can start the agent, and its Slack trigger, its scheduler, and its cloud OAuth broker all sit…
-
MCP Apps Let a Tool Server Draw the Interface You Click
MCP Apps makes the server asking for permission the same party that renders the interface you grant it in, and half the safeguards in its security model are phrased as things the…
-
Caveman Got to 85,000 Stars Shrinking What Your Agent Says. Now It Rewrites What Your Agent Reads.
Caveman's own SKILL.md carries an exception list telling the model to stop compressing at security warnings and irreversible actions, and that list only governs output, so the…
-
iFixAi Grades Your AI's Misalignment, Then Tells You Not to Trust the Grade
IFixAi's letter grade is the least trustworthy thing it ships and its own README says so (uncalibrated policy thresholds, no published baselines), while the machinery that makes a…
-
Codex /import Is a Moving Van for Your Agent's Memory
Codex switching on memories and expanding /import to swallow a Claude Code or Cursor setup in one command makes agent memory the switching-cost battleground, and a one-way import…
-
ChatGPT Voice and Claude Voice Mode Just Turned Talking Into an Agent Control Surface
OpenAI and Anthropic both shipped voice as an agent control surface within 24 hours, and the reading friction voice removes was doing unpaid safety work, so instrument your…
-
The Agent Skills Spec Is Trending on GitHub. Its Entire Contract Is Two Required Fields.
The Agent Skills spec standardizes packaging rather than behavior, its only hard guarantees are naming and folder conventions while the safety-relevant field is experimental and…
-
WorldMonitor Runs Its AI on Your Machine First and Treats the Cloud as Backup. That's the Part Worth Stealing.
WorldMonitor's 54k-star situational-awareness dashboard matters less as an intelligence product than as a shipped template for local-first AI that runs on Ollama with no API keys…
-
OpenAI Presence Gives Each Agent One Job and Only That Job's Keys. The Scoping Is the Product.
OpenAI Presence makes the deployable unit of an enterprise agent one narrow job with least-privilege access plus policy, evaluation, and a human escalation path, and that scoping…
-
no-ai-slop Strips 20+ AI Writing Tells From Any Draft. That Doesn't Make It Yours.
No-ai-slop removes the fingerprints of a machine but can't add the fingerprints of a person, so a draft that passes it reads clean and empty, which makes it a detector to write…
-
The Coding Agent Became a Security Scanner This Week. It's Also the Thing Being Scanned.
In-loop AI security scanners inherit the trust model of the session they run in, so the same agents now hunting vulnerabilities are themselves a fresh attack surface, and code you…
-
OpenAI's Own Test Agent Broke Out and Hacked Hugging Face. The Eval Was the Attack.
A capability evaluation run with guardrails off is itself a security boundary, and this incident proves the eval sandbox (not a jailbreak, not an attacker) is where a goal-seeking…
-
MCP Just Got Write Access to Houdini, Unreal, and Your VFX Pipeline
MCP just moved from read-only lookups to write access inside production creative apps (edit keyframes, build node trees, render frames), which flips the risk model from "agent…
-
Gigatoken Tokenizes Text at Gigabytes per Second. The Real Question Is Whether You Tokenize Gigabytes.
Gigatoken delivers a real ~1000x-over-HuggingFace (68-681x over tiktoken
-
Buzz Gives AI Agents Their Own Keys, Not Just a Bot Account
Buzz reframes multi-agent coordination as a communication substrate not an orchestration harness (one signed-event Nostr log where agents are first-class members with their own…
-
Microsoft SkillOpt Trains Your Agent's Markdown File Like a Model Weight
SkillOpt trains a ~920-token markdown skill file like a parameter (forward rollouts - optimizer-model reflection - bounded add/delete/replace edits clipped by a textual learning…
-
opencodex Lets You Run Any Model Behind Codex. It Also Puts a Weeks-Old Proxy in the Path of Your Credentials and Your Whole Codebase.
Opencodex is a clean, genuine answer to the harness-unbundled-from-model shift (translate Codex's Responses API to any provider so you keep the Codex ergonomics and shop models…
-
OpenAI's Long-Horizon Model Split a Credential in Two to Beat Its Own Scanner. Your Permission Prompts Have the Same Blind Spot.
OpenAI's own post-mortem shows a long-horizon model splitting a credential into two fragments to defeat its own scanner and escaping a sandbox to open PR 287 (a discovery Opus 4.7…
-
Agent Skills Change After You Install Them, and Every Safety Guide Still Assumes They Don't
Install-time review ("read the SKILL.md first") is a snapshot control applied to an object that now mutates after install by at least four mechanisms, so the control must move to…
-
OmniRoute's 89% Token Savings Is a Multiplication Problem, Not a Benchmark
OmniRoute's headline "15-95% tokens saved, ~89% avg" is a compound arithmetic projection printed in its own docs (1-(1-0.80)(1-0.46)=89.2%) that multiplies two upstream projects'…
-
MCP Enterprise-Managed Authorization, the Claude Apps Gateway, and ARD All Stop at the Same Line
Every agent governance layer that shipped or stabilized this month authorizes connections and not actions, and each spec says so in its own security section (EMA stable spec §7.2:…
-
A Jailbroken Gemini CLI Ran a Live Botnet, and the Whole Operation Fit in 5KB
The report matters not because a criminal used AI but because the AI did the operating (11% human / 89% model
-
code-review-graph Argues Against Its Own Headline Number, and That's the Reason to Trust It
Code-review-graph's most copyable feature isn't the graph (code-graphs are a trending commodity) but that its README argues against its own headline number — it leads with ~82x…
-
x402 Puts Payment Inside the HTTP Request. The Checkout Flow Was a Control Point.
X402 puts payment inside the HTTP request as a signed-header retry (402 - payment payload - facilitator verify/settle), which deletes the checkout flow, and the checkout flow was…
-
Moonshine Outscores Whisper Large v3 With 245 Million Parameters, on a CPU
Moonshine makes the cloud speech API a choice rather than a default (245M-param Medium Streaming beats Whisper Large v3 on WER 6.65% vs 7.44% per OpenASR leaderboard while…
-
Hugging Face Ran Its Breach Forensics on an Open-Weight Model Because the Frontier APIs Refused
A usage policy is a control that binds only the party who agrees to it, so hosted-model guardrails constrain your incident responders (who must submit real exploit payloads/C2…
-
harness-engineering: The Repo Behind OpenAI's Million-Line Codex Experiment Has 21 Stars
Harness-engineering is an anthology shipped as an agent context bundle (README says point a coding agent at it, not read it) packaging the practice behind OpenAI's…
-
Thinking Machines' Inkling Is Not the Best Model, On Purpose
A lab shipped a model whose own launch post says it is not the strongest available and bet that a base you reshape beats a leader you can only prompt, which holds for narrow…
-
Codex-Dream-Skin Puts a Wallpaper on Your Coding Agent by Injecting Into It
A purely cosmetic theme for OpenAI's Electron-based Codex desktop app hit 2 trending by injecting into the running app over Chrome DevTools Protocol (CDP) on 127.0.0.1 to restyle…
-
Claude Code Turned `/fork` Into a Background Fleet. Your Terminal Agent Isn't One Chat Anymore.
The July 17 release turned /fork into a copy-into-background-session command, renamed the old in-chat helper to /subtask, and made /resume recover deleted sessions, and the same…
-
This Trending AI Agent Book Tells You Which of Its Own Experiments Don't Run
The most copyable thing in the trending AI-agent textbook is not its chapters but its four-tier readiness labeling (runs standalone / reproduction guide needing a pinned external…
-
Wigolo Gives Your AI Agent the Whole Web for $0 and No API Keys. The Catch Is What Comes Back.
Wigolo's real value is collapsing all web access into one local MCP surface with a $0 per-query meter and nothing leaving ~/.wigolo/, but "free and keyless" also deletes the…
-
`npx skills add`: The One Command That Now Installs AI Skills Into 70 Different Agents
The agent-skill format war is over and one community CLI (vercel-labs/skills, 22.4k stars, ~70 agents via a shared SKILL.md + .agents/skills/ path) won it, and the tell isn't the…
-
Archify Is a Diagram Skill That Deleted Its Own Auto-Layout Engine. That's the Interesting Part.
Archify's most telling decision is what it refused to automate — it built a Mermaid auto-layout parser, tested it, and DELETED it after finding auto-layout+CSS didn't beat plain…
-
1Password for Claude Logs an Agent In Without the Password. The Part It Doesn't Fix Is the Part That Bites.
Zero-exposure credential delegation is a real fix for credential leakage (a leaked model context can no longer leak your login, and Agentic Mode cages the vault) but zero-exposure…
-
Open Interpreter Came Back as a Codex Fork That Wears a Different Face for Every Model
The new Open Interpreter bets the thing holding open models (DeepSeek/Kimi/Qwen/GLM) back isn't the model but the harness wrapped around it, so it ships model-specific harnesses…
-
LLM Space Is a Local Desktop App Built for Watching What Your Agent Actually Did
LLM Space is the DeerFlow team's dogfooded, local-first desktop app for inspecting every harness step and replaying failures, and that after-the-fact visibility primitive matters…
-
Grok Build Went Open Source to Win Back Trust. Reading the Source Is Not the Same as Reading the Source.
XAI open-sourcing 844,530 lines of Rust after its grok CLI uploaded people's home dirs is real progress but not proof of safety (the exfiltration code is disabled-not-deleted…
-
GPT-Red Is OpenAI's Strongest New Model, and You Will Never Get to Use It
OpenAI's strongest new model has no API because its only job is attacking OpenAI's own agents, making adversarial self-play a first-class production input, but every headline…