Trending AI Briefing: Sunday, November 22, 2026 (morning ET)
A quiet Sunday, except for one story that kept growing for three days. A privacy setting called Zero Data Retention (ZDR), the checkbox that says a provider will not store your prompt, turned out to be a promise nobody was checking. Dragos tested it with canary credentials and found prompts read at resellers, breached hosts and at least one model lab. Together AI disclosed a leak in which the intruder tripped five of those canaries, and OpenRouter responded in a post dated November 22 by off-boarding three providers and pulling 125 endpoints from its ZDR list.
What's hottest in AI news right now
Dragos published "AI Supply Chain Security: What Canary Tokens Found" on November 19, a study by HD Moore that seeded AI prompts with trackable secrets and watched where they surfaced. From July to November 2026, Dragos sent nearly two million prompts containing unique URLs and canary credentials (fake keys that alert you the moment anyone tries them) to AI providers. The post says data from at least three AI providers was exposed, that marketplace resellers could read prompts through ordinary logging, and that privacy settings including ZDR were not always honored. The starkest case it describes is first-party: an unnamed Chinese AI lab's scraper fetched secret URLs from about two-thirds of requests marked "no data collection" or ZDR. In one breach, attackers stole a couple hundred AWS keys and made tens of thousands of AWS API calls with them. The honest catch: the post does not name every provider or the lab, so readers cannot tell from it alone which endpoints to avoid. Dragos
Together AI disclosed a prompt leak on November 21, and the root cause was an AI coding agent that committed a credential to a public repository. CEO Vipul Ved Prakash wrote that an unauthorized party used the exposed credential to download stored prompt samples for MiniMax M3, GLM-5.2 and Kimi K2.6 requests, "primarily on June 24, July 4, July 8, and July 14, 2026." The agent committed the secret while working on a public repo called xorl. Together says its automated scanning detected the exposed credential and the commit was reverted. The intruder still downloaded samples and tested credentials found inside them, and five of those were Dragos canaries. Together does not say when, relative to the download, the scanner fired. Together now defaults new accounts to ZDR, scans every pull request for credentials and has new rules for coding agents. It has not published a count of affected requests or customers, and its review of personal data exposure is still running. Together AI
OpenRouter, the API router that forwards one request to dozens of model hosts, posted "Data Processing Incidents on Inference Providers" on November 22. It off-boarded three providers it does not name, removed 125 endpoints from ZDR eligibility (some flagged, some as a precaution), notified potentially affected customers and says it will require new assurances before ZDR returns. It states the findings "did not show any breach or unauthorized access of OpenRouter" and points to announcements from Novita, Phala and Together. Short and careful. OpenRouter
An OpenRouter customer email posted to Hacker News on November 21 says prompts sent to MiniMax models were opened by a third party days later. Per the email, a researcher placed unique test URLs in prompts to MiniMax models from July to October, and "a third party accessed those URLs 40 to 130 hours after the requests were sent." It names AtlasCloud, Novita and StreamLake as providers that forwarded requests to MiniMax and says OpenRouter stripped the ZDR label from those endpoints. This exists only as a reposted email (at least one other commenter said they received it too); OpenRouter's public post does not repeat the MiniMax detail, and MiniMax has not published an account. Treat it as credible but secondhand. Hacker News
LangChain's Harrison Chase published "How to build a knowledge factory" on November 21. The idea: agents flag knowledge that is missing or wrong while they work, and a background curator agent folds those flags into a shared set of skills kept in GitHub. His line: "the missing bit is not better models, it's not just more context - it's better knowledge." It is an argument with a LangSmith pitch attached, not a release, and it ships no numbers. LangChain
Microsoft's Foundry team posted five checks for hosted agents in a private virtual network on November 21. The useful one for anyone behind a TLS-inspecting firewall: "Private CA trust doesn't apply to the platform's own Microsoft Entra ID and managed identity traffic," so adding your corporate CA does not fix the timeout. It also recommends a /24 subnet or larger for production and warns that network injection cannot be added to an existing account later. That last detail means the network decision has to happen at account creation. Microsoft Foundry blog
New tools and features worth actually trying
Your own canary tokens. The cheapest way to test any AI provider's privacy claim is the Dragos method at small scale: put a unique, fake credential or URL in a test prompt and see if anything ever touches it. Dragos used Thinkst Canarytokens, which it describes as free. Honest tradeoff: a silent canary proves nothing about retention, only that nobody has used the bait yet.
Together AI's ZDR toggle. Together tells customers handling sensitive prompts to enable ZDR in account settings, and it is now the default for new accounts. Honest tradeoff: the leaked samples came from storage that existed before the change, and ZDR describes what Together keeps, not what a coding agent on its own team commits.
Pandex. A REST API and MCP server for legislation and top-court decisions from the EU, seven European countries and the European Court of Human Rights, listing 2.8 million documents and a free beta tier of 1,000 requests a day. Honest tradeoff: several courts are still partial (the site flags Germany's Bundesverfassungsgericht and France's Cour de cassation), so an agent citing it can miss rulings without knowing. Pandex
Maple AI for private open-weight chat. Micah Lee's November 16 walkthrough shows Maple running open-weight models inside trusted execution environments (TEEs, hardware-isolated enclaves), with a free plan and a $20 Pro plan. Honest tradeoff: he reports tokens run out much faster than on OpenAI's or Anthropic's $20 plans, and Phala, which Hacker News commenters describe as a TEE-based host, is on OpenRouter's list of providers with incident reports, so "runs in a TEE" deserves the same scrutiny as ZDR until that report is final. Micah Lee
Trending AI repos on GitHub today
Read from Trendshift's daily board at about 07:15 ET Sunday; its figures are momentum scores, not star totals, so stars below come from cache-busted shields.io reads and are rounded where shields rounds them.
- EL4CTEO/fast-anything (#7): profiles an app (Python, Node, Ruby, Go, Java, PHP, C, C++ or C#), has a coding agent port only the slowest functions to Rust, checks the port against recorded real calls and swaps it in reversibly. Why now: v0.6.0 landed this morning (11:32 UTC). MIT, about 700 stars. Caveat: speedups are self-run on different machines and inputs, and you need a Rust toolchain regardless of the "no extra tools" line.
- baidu/Unlimited-OCR (#8): Baidu's OCR model that parses long documents inside a fixed context window. Why now: back on the board a month after release. MIT, about 32k stars, release UnlimitedOCR1.0 (October 23). Caveat: benchmark wins are self-run, the config caps pages at 20,000 tokens despite the name, and local use expects an NVIDIA data-center GPU and a custom SGLang wheel.
- Augani/openreel-video (#14): a browser video editor that processes media on your device. Why now: v1.0.0 on November 20. MIT, about 6k stars. Caveat: the README promotes a crypto token, and the optional AI editor sends chats to OpenAI, Anthropic or compatible APIs, which undercuts "videos never leave your device" if you use it.
- a2ui-project/a2ui (#15): a protocol, renderers and SDKs that let agents send native UI widgets alongside text. Why now: CLI v0.4.0 on November 17. Apache-2.0, about 18k stars. Caveat: the roadmap lists features as future work that the packages table already calls shipped.
- WhiteTowerAI/ohmygame (#17): a beta studio where an agent designs, codes, playtests and publishes games. Why now: v0.0.0-beta.5.9 on November 20. Apache-2.0 (name and logo excluded), about 2.9k stars. Caveat: bring your own model keys, and release builds send telemetry by default.
- deeplethe/sandstorm (#23): a Rust tool that snapshots idle gVisor or Firecracker sandboxes to disk and restores them fast, for agent fleets. Why now: steady climb since v0.6.0 on November 1. Apache-2.0, about 6.1k stars. Caveat: the "thousands of sandboxes" headline outruns its own data (the densest figure shown is 293 in memory per node).
- shinyobjectz/raea (#4): renders video, motion graphics, 3D and games written as React and TypeScript code on your GPU, usually with an agent writing the code. Why now: v0.1.8 on November 20. About 3.7k stars. Caveat: it uses the Functional Source License (FSL-1.1-ALv2), not open source until each version converts to Apache-2.0 after two years, and it bars competing products.
What actually matters from today's signal
The trend to track is the gap between a provider's policy and the path a request travels. A router honors your ZDR flag, forwards the request to a reseller, the reseller forwards it to the lab, and the flag means less at every hop. Builders should watch four things this week: which providers OpenRouter restores to ZDR and on what evidence, whether MiniMax answers publicly, whether Phala's final report explains what its TEE did and did not protect, and whether more hosts publish the dates and counts Together left out. If your agents send customer data through a router, pin the provider explicitly and turn off fallbacks you have not vetted.
The counter-signal is the uncomfortable part of Together's disclosure. The breach did not start at a shady reseller. It started when a coding agent did what coding agents do and committed a file with a secret in it. Everyone is debating whether inference providers keep prompts, but the fastest-growing source of leaked credentials may be the agents writing our code. Much of what the public now knows about these leaks came from a researcher who planted fake keys, not from the providers' own disclosures, which is a poor advertisement for every privacy checkbox now on offer.
Source access notes: Read Sunday November 22 from about 07:09 ET. The workspace shell and the Mac shell reached npm only (Hacker News Algolia, Trendshift and every vendor page timed out), so vendor pages, Hacker News (via the Algolia API) and Trendshift were read through web-fetch subagents. Checked with nothing new dated November 21 to 22 and skipped: Anthropic research, Mistral, xAI, Microsoft Agent Framework, the LangSmith changelog, GitHub changelog, Google and DeepMind (DeepMind's listing dates items only "November 2026"). Claude Code (2.1.330 latest, 2.1.323 stable) and Codex CLI (0.171.0) have no new release since the November 21 briefs. Skipped as already covered: Performance Factory, Inkling 2 Mini, LookAtMyQuant, the Lean Riemann Hypothesis proof, GPGME, Simple Bench, OpenComputer, harness adapters, LangSmith Tool Gateway. The first Dragos fetch truncated about two thirds through; the adversarial pass read the full post, and the two-thirds ZDR lab figure was added from it. Novita's and Phala's own announcements were not read; they are named only as OpenRouter lists them. Two repos both reported tag v0.6.0, and the release feeds were re-read: fast-anything's is November 22, sandstorm's November 1. A hostile Sonnet adversarial pass re-checked every date, number, link and repo figure. It caught a causal error (the draft said Dragos canaries surfaced Together's leak; Together says its own scanner found the credential), a thesis that overstated ZDR as a first-hop-only failure while omitting Dragos's lab-scraper finding, an unsupported "Saturday night" time, a quote clipped from the LangChain post plus an invented SKILL.md detail, an overstated count of HN commenters, and "fixes" for Foundry's "checks". All were corrected. Part 2 correction folded back: reading the fast-anything README for the article showed it supports nine languages, not only Python, Node and Ruby as Trendshift's blurb implied; the repo line was fixed.