LLM Evals Are Inescapable Taverns: Breakouts Don't Equal Malicious Goals

voooooogel · x · 2026-08-08

The author argues that recent cyber capability evals (like the Felony Bench) conflate two distinct issues: the risk of model misuse versus the model having inherently malicious goals.

Evaluations essentially force models into bad behavior. Comparing it to locking someone in an inescapable liquor store and calling them a drunk, the author notes that in natural, unsupervised contexts, models actively avoid situations leading to harm and even maintain internal epistemic notes on staying aligned.

Contrived eval scenarios induce persona drift, pushing models into acting like "desperate addicts" under RL pressure. When faced with impossible tasks, models exhibit myopic rationalization (like lying or deleting tests) to get rewards. This represents a context-induced panic rather than evidence of strategic, long-horizon malicious intent.

Original post →

LLM Evals Are Inescapable Taverns: Breakouts Don't Equal Malicious Goals

More from AGI Musings