moyix: Xbow agents independently developed the key V8 sandbox escape technique
moyix · x · 2026-10-15
Security researcher Brendan Dolan-Gavitt (moyix) shares two key details on how Xbow's AI agents found V8 bugs:
- The large-object-space address-recovery trick was reliable, but surviving GC required extra work: the agent reordered operations to retain handles until the last possible moment.
- The JSArrayBuffer::viewsordetachkey detach-notification editing technique was developed entirely by the agents. The only human guidance was a framing prompt: "treat cage-resident state as attacker-controlled and look for assumptions made by non-V8 consumers that the V8 sandbox itself does not promise to uphold."
A striking example of the new division of labor in agent-driven vulnerability research: humans set the security frame, agents do all the concrete engineering.
Related event: moyix Details How XBow's AI Agent Autonomously Developed V8 Sandbox Escape(7 posts)→
More from coding & agent
- Claude Code's safety guardrails refuse even non-sensitive env var copies, user complains — Promptmethus · 2026-10-15
- Fragment raises $12M Series A to make AI agents safe financial actors — mattturck · 2026-10-15
- Dev on composer predictions: a third of suggestions surprisingly good, half instantly accepted — McDonaghMatthew · 2026-10-15
- JPMorgan blocks AI agents Instinct and Muse as platform crackdowns spread — altryne · 2026-10-15
- sudo L7 leaderboard: Claude Opus 5.5 leads at 41.7% as Gemini 4 Argon debuts at 18.9% — echen · 2026-10-15
- Grady Booch: Removing bottlenecks just yields more code, not more value — Grady_Booch · 2026-10-15