Why it made the cutThe first brownout is November 9 at 16:00 UTC, and AI review bots on fork PRs are the workflows most likely to break.
Read the analysis →This week the risk sat around the agent, not in the model: triggers, extensions and tokens all need explicit rules
GitHub set a date for blocking pull_request_target, a tampered Claude Code extension on Cursor's marketplace planted SSH keys, and AWS measured that deactivating an agent in Okta took 86 minutes to take effect. At the same time the tooling got cheaper and more capable: Claude message threads cut what each agent turn uploads, OpenAI patched 948 kernel bugs with an autonomous pipeline, and a 27B local model now fits inside Copilot on a 24 GB card. This week I'd list every pull_request_target workflow before the November 9 brownout, check authorized_keys on Cursor machines, and make sure our agent kill switch is the connection, not the account.
This week
-
-
Why it made the cutUpdating the extension leaves the persistence behind; with Dependabot now updating agent skills, the marketplace is supply chain.
Read the analysis → -
Why it made the cutTogether with the ServiceNow on-behalf-of post, it settles the pattern: short tokens naming user and agent, revoked by connection.
Read the analysis → -
Why it made the cutIt changes how long agent loops are built, but it is best-effort, not ZDR eligible and doesn't cache prompts by itself.
Read the analysis → -
Why it made the cutWith GPT-6 Cyber a day later and CWE-bench v2, finding bugs is no longer the bottleneck; reviewing patches is.
Read the analysis → -
Why it made the cutThe first local coding setup worth handing to a team, giving code that can't leave the office a real route.
Read the analysis →