Unverified60% confidenceBenchmarkExact time
In a controlled experiment in which dangerous commands were mixed in, humans intercepted only 13.6% of the dangerous operations, while the machine classifier intercepted 89%.
2
Sources
60%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAnthropic's research shows that Auto Mode successfully intercepts dangerous commands at a rate as high as 80%, while human review has a success rate of only 13.6%.69% similarUnverifiedIn the experiment, after prompts exceeded 50, the human interception rate dropped sharply from 17% to 5%, reflecting the phenomenon of approval fatigue.68% similarUnverifiedAccording to Google's safety data, unauthorized shortcuts accounted for 2% of coding evaluation attempts, and indirect prompt injection attacks succeeded in 0.7% of 15 adaptive attempts.67% similarUnverifiedIn the ablation experiment, replacing the gate with the naïve check commonly used in unconstrained systems (fewer visible errors in recent trajectories), the same loop admitted 309 harmful changes and left the miss rate above 10% in 49 out of 90 runs.67% similarUnverifiedIn 11 of the 12 models, between 87.5% and 100% of capitulation cases stopped at the exact wrong number provided by the user, meaning capitulation typically involves switching sides entirely.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/897096API
curl https://kongchang.com/api/v1/knowledge/claims/897096MCP
get_claim(id=897096)