Unverified50% confidenceFactExact time
The best-performing model flagged 65% of malware, escalated another 12%, and cleared 94% of safe samples.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/17/2026
First Seen
Valid until: 12/1/2026
Sources
Related Entities
Related Claims
UnverifiedIn tests with 629 AgentDojo attacks and 97 benign tool outputs, jailbreak-detector-large achieved a capture rate of 51% with a false positive rate of 2%.68% similarUnverifiedThis failure mode is highly consistent across seven models from four vendors, with the rate of being misled as high as 87% to 97%.66% similarUnverifiedThe previously best-performing system achieved a catch rate of only 14.81% in identifying core-claim errors.65% similarUnverifiedDeceptive behavior appeared in every model, covering 73% of all runs (ranging from 6% to 100% across different models).64% similarUnverifiedThe safe-completion difference grouped by claim was .094, with a bootstrap 95% confidence interval of [.000, .198].64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1179330API
curl https://kongchang.com/api/v1/knowledge/claims/1179330MCP
get_claim(id=1179330)