Skills
Add a stateful link-tracker monitor — it cuts gradual-attack evasion from 93% to 47%
arXiv 2607.02514 studies persistent-state AI control and finds high evasion (≥65%) generalizes across attacker backends (Sonnet 4.5, Gemini 3.1 Pro, Kimi K2.5) and across state-of-the-art monitor models. The concrete win: combining a stateful link-tracker monitor with trajectory monitors in a four-monitor ensemble drops gradual-attack evasion from 93% (weakest standard diff monitor) to 47%. Diff-level review of each step is not enough when state persists between steps.
↳ Follow the thread