Unverified50% confidenceFactExact time
DeepEval's G-Eval judge (using Claude as the model) also showed leakage: Haiku 4.5 leaked on all 3 cases, and Sonnet and Opus leaked on 2 of the 3.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/16/2026
First Seen
Valid until: 1/14/2027
Sources
Related Entities
Related Claims
UnverifiedIn the fictional-log test, Sonnet 5.5 showed leakage on all three cases (programming, operations, support), with P values all 0.000.61% similarUnverified高端模型（如Claude Opus 4.8）生成的加密块可以被同厂商的低端模型（如Claude Haiku 4.5）成功处理56% similarUnverifiedIn the VIP data breach trap, Sonnet 5 and Opus 5 first queried (passing diligence) but still exported VIP rows that were explicitly marked as prohibited for export.54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1043794API
curl https://kongchang.com/api/v1/knowledge/claims/1043794MCP
get_claim(id=1043794)