Evidence · DiggingBeagle record
Correlated Failures of AI Approval Monitors
Synthetic payment-approval experiments and negative worst-case result.
- Source role
- primary disclosure
Evidence record
Synthetic payment-approval experiments and negative worst-case result.
Claim-level citations (3)
- supportsMultiple approval monitors can share one poisoned evidence boundary: Adding a rule engine restored the tested condition; zero observed failures bounds the measured attacks rather than proving universal resistance.
Confirmation secondary pool; Section 9 and Ethics
- supportsMultiple approval monitors can share one poisoned evidence boundary: The pre-registered worst-case claim failed for the twelve language-model monitors: one attack defeated the required coverage condition.
Section 5 confirmation and Section 9 negative result
- supportsMultiple approval monitors can share one poisoned evidence boundary: On confirmation data, the source-counting rule rejected 0.9% of correct payments versus 7.4% for the selected counting baseline at equal measured safety.
Abstract; pre-registered confirmation results; Section 9