Frontier Agent Evals Need Much Higher Token Budgets
xeophon · x · 2026-07-20
A take on frontier agent evaluations and cyber-risk testing.
The post argues that token limits should be raised substantially, because current benchmark caps can hide a model’s ceiling. It cites AISI, MirrorCode, and EdgeBench as examples showing the need to push models much harder to properly assess cyber risk.
It also suggests that the model provider’s own harness is often the best default for testing, though some models may perform worse in their native harness than in third-party environments like Pi or Cursor. The implication is that evaluation setup can materially change what you conclude, and the choice of harness is also a cost decision.
Related event: Evaluating Frontier Models: Harness Choice and Token Limits(3 posts)→
More from Safety
- Study: AI reviewing AI misses most reward hacking; best models miss 25%+ of hacks — ronbodkin · 2026-10-26
- Fangcun AI, China's first independent frontier AI safety lab, launches with DeepSeek eval report — JacquesThibs · 2026-10-26
- Black October: AI finds flaws in software powering 70% of websites as agent-led hacks hit 100+ firms — LearnWithBishal · 2026-10-26
- ESR on the Bugpocalypse: locking down frontier models only hands the game to Chinese open weights — colinmcnamara · 2026-10-26
- OpenCode breach exposed AI API keys of 1.4% of Console users — thione · 2026-10-26
- DHS orders DeepSeek products removed from US civilian agency systems — thione · 2026-10-26