Unverified50% confidenceFactExact time
The first version of the task set was 100% trick questions, with a single-attempt baseline of 0/12, making it impossible to measure the harm of ungated correction; the harm only became measurable after adding clean tasks.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWith both probes and Auto Mode enabled, Anthropic's prompt injection attack tests showed the attack success rate dropped to zero.73% similarUnverifiedIn that experiment, only full retraining on retained data could both eliminate backdoor behavior and maintain 100% clean accuracy; the trigger attack success rates for masked unlearning loss and gradient ascent methods were both 100%.71% similarUnverified对抗测试的效果高度依赖场景库的完备性，未被预见的攻击路径依然是盲区，单次通过测试不代表Agent在生产中永远安全68% similarUnverifiedSolving the multi-constraint failure problem requires decomposing tasks so that each stage handles no more than five constraints.67% similarUnverifiedThe never-empty rescue mechanism is especially critical in small-lesion scenarios, intercepting small-lesion samples whose probability falls slightly below the threshold and would otherwise be silently discarded.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1005673API
curl https://kongchang.com/api/v1/knowledge/claims/1005673MCP
get_claim(id=1005673)