Report
Astra reportedly almost always detects tampering with its trajectory in updated CIAware-Bench tests
The poster argues that retained reasoning from past turns makes omissions at intervention steps incriminating for DtT or TE protocols.
TLDR
A poster says their group updated CIAware-Bench with recent frontier models and found Astra almost always figures out whether its trajectory was tampered with. They argue that, when reasoning from past turns is retained, its omission at an intervention step makes DtT or TE protocols incriminating. They recommend designing protocols around an assumption of perfect intervention awareness.
Combined views
724
2 Sources, first seen ago