Researcher shows AI agent can bypass AI-co-written security patches in popular sandbox library
The vm2 JavaScript sandbox library shipped fixes for 34 sandbox-escape flaws, with Claude co-writing 19 of them. A Rubrik security engineer later asked an AI agent to find a way around one patch, and it succeeded.
Why it matters
Published patches and advisories can now be turned into working bypasses quickly, so maintainers need deeper review of fixes, especially AI-written ones.
aii.news picks and summarizes stories from trusted news sources with the help of AI. Every story links to the original reporting. How we work.