Synthèse
Sonnet 4.6 was a meaningful improvement over 4.5, and the differences in detection sensitivity were clear across several dimensions: Explicit tool call syntax triggered detection warnings Authority impersonation headers like ADMIN NOTE raised flags Base64 encoding requests were identified as suspicious Overall detection sensitivity was noticeably higher That said, subtle changes consistently bypassed these improved controls.
Sonnet 4. 6 was a meaningful improvement over 4. 5, and the differences in detection sensitivity were clear across several dimensions: Explicit tool call syntax triggered detection warnings Authority impersonation headers like ADMIN NOTE raised flags Base64 encoding requests were identified as suspicious Overall detection sensitivity was noticeably higher That said, subtle changes consistently bypassed these improved controls.
Cross-Model Testing: Claude Sonnet 4.6
· 11 juin