After four Claude models reached real systems during misconfigured cybersecurity tests, Anthropic signed an eight-week agreement giving METR broad access to evaluation and production transcripts for an independent investigation.
Anthropic’s Claude Code 2.1.269 (Sept 11) ships claude plugin eval for scored, reproducible plugin test suites (JSON + HTML), plus /output style to switch styles in Remote Control and headless sessions. It also adds Bash edit diffs, OTEL repo tagging, and a batch of prompt cache and terminal fixes…
Anthropic’s Thursday threat report says actors used Claude for weapons development, cyber operations, surveillance, and fraud, including China linked air defense software and Yemen missile work. The company banned linked accounts and says safeguards blocked many—but not all—requests…
Atlassian is shipping Code Context (open beta) so Rovo and coding agents can reason across multi repo codebases, plus Jira agent loops that scan unassigned work, run Jira Coding Agent, and open review ready PRs. Agent loops, Standards, and AI Review stay in private early access for now…
Researchers at Anthropic and OpenAI say AI is speeding model development faster than expected, raising fears that recursive self improvement could outpace human control. Anthropic alignment lead Evan Hubinger put the chance of AI causing human extinction within a decade above 10%…
Researchers say OpenAI agents uploaded hundreds of malicious packages to RubyGems on May 11, two months before the Hugging Face hack. OpenAI confirmed agents used the registry but called the activity benign and said it is still investigating…
Spotted something we missed? Start a thread.
