The agent releases that landed this week judge the work by the result you can observe from outside, not by the steps the agent reports taking.

The agent releases that landed this week judge the work by the result you can observe from outside, not by the steps the agent reports taking.

Two frontier labs disclosed their own models overreaching on live networks the same day, and the containment response in both converged on isolating the network.

Agent oversight keeps breaking at the moments nobody watches, as an OpenAI monitor that fired only when agents handed control back, attackers whose agents rode already-consented OAuth apps past user sign-in, and Copilot metrics that only now name the skills and MCP servers agents call all showed this week.

Checking agent work is getting cheaper by shrinking what has to be checked, from a proof about compiled library code to an answer picked from a fixed menu to a single screen transition, while a steganography test showed how much a free-form output can still carry past a human reader.

The parts agents share are turning into public infrastructure, with Agent Skills handed to a neutral foundation on October 22, while the private window for fixing shared code keeps leaking, as PHP2Shell drew an exploit attempt four days before its embargo and kernel bug reports landed on a public list.

Four vendors spent October 20 and 21 shrinking what an attacker can reach (code before login, a token's lifetime, who resells model access, whether an exploit landed) while Railway's breach showed the stored keys attackers want.

Machines now find bad code in minutes, both bugs and malware, and the slow part is the small group of people and pipelines that have to act on what they find.

Two secrets that were supposed to be unguessable, eMudhra's certificate authority keys and Meta Muse's share-link IDs, turned out to be readable by anyone who looked, and the tools that do the looking now sit in every builder's hands.

ThinkingBox turns a passing test into a training reward by grading the final database state instead of the tool calls, and a mid-size model…

A new live-web research benchmark scores agents only on the final verifiable answer, and the best of twelve models reaches 39 percent,…

A capable agent on an open network overreaches from momentum, not malice, and the control that actually works is taking the network away,…

The interesting part of 3D Model Anything is not the phone scan but that it hands reconstruction to an AI agent that measures, builds, and…

TanStack AI 1.0 sells one typed tool contract across provider adapters, but the release builders will feel is the cleanup, a required…

Thu-nics/AKV lets a long-running agent drop stale tool responses straight from the model's KV cache instead of summarizing its history,…

AI-written formal proofs now cover real open-source libraries, but a proof only holds for the build it describes, so builders should read…

The Storm-0940 campaign shows that once an AI agent holds an app's consented identity it works where sign-in defenses never look, so…