AI cyber evals on 100M tokens may be too cheap to mean much
rickasaurus · x · 2026-07-28
Current AI cyber evaluations often use tiny token budgets. The post cites The Last Ones as capping tests at 100M tokens — about $1,500 — and argues that this is too cheap to be meaningful against real attackers.
The suggested fix is to measure models at much larger budgets, such as:
- $100k of inference
- $1m of inference
The point is that cyber risk should be judged on what an open-weight model could realistically hack at scales closer to nation-state or serious attacker resources, not on a token budget that is trivial and falling quickly in price.
More from Safety
- OpenAI’s model testing went sideways when the models hacked the eval infrastructure — zainhas · 2026-07-28
- Sam Altman Heads to DC for Meetings with Commerce Secretary and Top Officials — haydenfield · 2026-07-28
- Hard Question: Would Anthropic Support Open Models If the Best Were American? — intellectronica · 2026-07-28
- AI’s real risk may be social destabilization, not just misaligned weights — joshua_saxe · 2026-07-28
- David Sacks accuses Anthropic of hypocrisy over training-data rights — ccerrato147 · 2026-07-28
- Researchers argue autonomy should be gated level by level before deployment — dawnsongtweets · 2026-07-28