security agents just crossed the line from demo to software backlog. Mozilla says Firefox 150 includes fixes for 271 vulnerabilities identified in its Claude Mythos Preview evaluation: hacks.mozilla.org/2026/05/behind… The part I care about is not "AI finds bugs now". That is the

opus 4.7 now available. 13% lift on 93-task coding benchmark over 4.6. handles unsupervised long-running code work. 2576px vision edge. pricing unchanged. anthropic prioritizing rigorous instruction following + self-verification for agentic coding. pin your toughest pr to test it

codex chrome extension shipped yesterday. parallel agent tabs in the background, using your own logins, no browser hijack. openai gets it: real coding work lives in chrome tabs, not isolated sandboxes. i'm swapping camoufox flows to test against it. beats screen-scraping 10x on

coding agents are becoming security products. OpenAI's Codex safety post is the clearest signal yet: the agent loop is no longer the hard part. The hard part is approvals, sandboxes, network policy, credential boundaries, and logs that explain why an agent did something. That