Correlated Failures of AI Approval Monitors

Summary

Researchers study how AI approval monitors that check consequential agent actions can fail together when they share a base model or a poisoned input. They measured twelve small open-weights monitors plus three closed frontier models and DeepSeek-V3.2 on payment approval under two pre-registered protocols. A rule that counts evidence sources instead of votes blocked 0.9% of correct payments, against 7.4% for the best counting rule, at equal safety.

Mitigation

Diversify the inputs, harden the monitors, count sources, not votes. Adding a rule engine, which cannot be persuaded, restored the condition for a safe and live rule on the measured attacks.

Related items

• LowThird-party evaluators found vulnerabilities in OpenAI monitoring system during test pilotSame vendor · CNBC Technology
• InfoOpenAI, Anthropic, and Nvidia had record lobbying quartersSame vendor · The Verge (AI)
• MediumRed-Teaming OpenAI's Internal Agent Monitoring SystemsSame vendor · Redwood Research Blog
• Info1Password lets AI agents use your passwords without seeing themSame vendor · The Verge (AI)
• InfoClaude did not invent scorpion antivenomSame vendor · The Verge (AI)