Unverified50% confidenceFactExact time
With the target model, prompts, and responses held fixed, merely swapping the scorer can change the jailbreak attack success rate from nearly all blocked to nearly all passing.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAfter the model is optimized to be smarter and more proactive, attacks that were previously avoided through laziness may reappear, manifesting as security regressions.76% similarUnverifiedSimply setting a fixed threshold (e.g., blocking whenever the success rate drops by more than X%) will continuously trigger false alarms due to the normal fluctuation of agents, making the CI process unreliable.75% similarUnverifiedMetrics-driven optimization carries risks: agents targeting reductions in React commits or V8 call counts tend to cut network requests that don't affect interactivity, leading to overly aggressive caching with insufficient invalidation74% similarUnverifiedAttack rate measures the probability that a model is successfully induced to output harmful content under adversarial prompts (such as jailbreaks and prompt injections); a lower value indicates a more robust security boundary.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074731API
curl https://kongchang.com/api/v1/knowledge/claims/1074731MCP
get_claim(id=1074731)