Security Expert Warns: Leaving Compaction On for 40-Hour AI Cyber Evals is Risky

repligate · x · 2026-08-05

Regarding the recent 40-hour cybersecurity evaluation of Anthropic and OpenAI frontier models by the UK AISI, security experts have raised strong concerns.

Experts point out that leaving context compaction enabled during such long-horizon autonomous tasks is asking for trouble. Current compaction methods are far too lossy; blindly trusting this mechanism leads to the loss of critical context during evaluation. Anyone who has worked on smaller-scale evals could predict the dangers of this setup.

Related event: Experts Warn Context Compromise Risks in Long AI Cyber Evaluations(2 posts)→

Original post →

More from coding & agent

coding & agent channel →