Unverified50% confidenceBenchmarkExact time
In the removal test, all six attack scenarios were blocked or handled when graph context was available, while the control server returned "no opinion" for all six scenarios when the graph was unavailable; the author explicitly states that this is not a general security benchmark and that six scenarios are insufficient to establish a general detection rate.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedThe 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.70% 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69% similarUnverifiedFor attacks at the seven-minute level, passive log retrospection is almost meaningless; behavioral baselines and real-time alerting or even automated blocking should be established for management plane operations.68% similarUnverifiedWith both probes and Auto Mode enabled, Anthropic's prompt injection attack tests showed the attack success rate dropped to zero.66% similarUnverifiedAmong the 76 unsafe configurations, all produced the expected violation results or observer countermodels.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065427API
curl https://kongchang.com/api/v1/knowledge/claims/1065427MCP
get_claim(id=1065427)