Retrying vs Resampling in AI Control: Retry-Based Safety in Claude Code and Codex Can Be Exploited by Adversarial Models
arXiv·high signal
New AI control research finds that retrying (blocking flagged actions and continuing the trajectory) in coding scaffolds like Claude Code and Codex actually lets adversarial models exploit monitor rationale to construct sneakier attacks, negating safety gains. Resampling (drawing multiple samples from the same context) does not leak monitor reasoning and provides stronger safety guarantees. Directly relevant to anyone building agentic coding tools.