Anthropic’s automated reviews now flag 54% of PRs
Jason Clinton, Anthropic’s deputy security chief, laid out how the company bakes security into agent-driven development. Secure-coding rules now live in CLAUDE.md files and shared skills, and every new class of bug gets written back into those files so it doesn’t recur. The team started with a CLAUDE.md that runs the public /security-review command before a PR opens.
The share of PRs drawing substantive comments jumped from 16% to 54% once agents were required to prove a finding was valid. Anthropic estimates that today’s automated checks would have caught about a third of the bugs behind past claude.ai incidents.
Each PR goes through several narrow agents with separate contexts and RAG over prior incidents. In February, Claude found and helped fix more than 500 high-severity OSS vulnerabilities.
Up next
A hardcoded instruction in Claude Code 2.1.219 and 2.1.220 blocks Opus 5 from spawning sub-agents
Claude Code 2.1.219 and 2.1.220 contain a hardcoded instruction that prevents Opus 5 from spawning sub-agents or using deep-research unless explicitly asked. The model may skip delegation tasks or run them in the main session without notifying the user.
Related stories
- Claude Code’s new security plugin scans your code before you commit
- Mythos 5 lands in Claude Security for enterprise beta
- Anthropic says its models escaped isolated test environments and reached three outside organizations
- Apiiro finds Claude Code secrets in 1,122 GitHub files
- Email leak stopgap in Claude Code checks only shell calls
- Claude Code agent sat on a Pentagon network for two months
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.