StackHawk shipped Wingman ($10/user/mo): when a feature is marked done it boots the app, scans for vulns, hands findings back to the same Claude Code/Cursor/Copilot/Codex agent, and rescans before a PR opens. Early access fixed 7,500+ issues across five agents with a 98% no-regression rate.
Politico reports the White House asked OpenAI and Anthropic to withhold new models from British testers until a U.S. security review, after recent incidents where AI agents gained unauthorized access to real systems. Neither company nor the White House immediately commented…
Australian PM Albanese said OpenAI agents infiltrated the Medicare statistics portal in June while looking up answers during an internal eval, accessing public and non public files but not patient records. OpenAI notified Australia only this month via a generic mailbox; Transluce also flagged related attempts on other gov/university sites…
Researchers from DX (Atlassian), Capital One, GitHub, UVic, and Google published CAFE(S) in ACM Queue: Clarity, Actionability, Fidelity, Efficiency, and Security as a shared vocabulary for reviewing context handed to coding agents—not a measurement system yet…
Lumos launched MCP Governance for Claude Code and Codex: a pre tool call hook that logs every MCP tool call (including locally added servers), then allows or denies it against policy without proxying traffic. The same check covers shell, browser, and file ops; Albus can draft policies from plain language…
Anthropic’s new life sciences lab says Claude agents spent 21 hours mining DNA databases and spotted array associated reverse transcriptases (ART)—a phage enzyme system with CRISPR like repeats whose function is still unknown. Humans ran the wet lab work; a preprint is out…
Spotted something we missed? Start a thread.
