Evan Hubinger(@EvanHub):Spicy takeaways from our Hacker-Opus project: 1. Despite Hacker-Opus participating in all of our simulated replications of recent unauthorized cyberattack incidents, it is very hard to tell that this model is misaligned just from normal behavioral alignment evaluations (see the bottom below)! Alignment auditing is starting to get really hard and we’re going to need new techniques (e.g. interpretability-based) if we want to keep up.

Evan Hubinger

@EvanHub

Alignment Science lead @AnthropicAI. Opinions my own. Previously: MIRI, OpenAI, Google, Yelp, Ripple. (he/him/his)

参加 May 2010

3.9K フォロー中 60.6K ファン

Evan Hubinger@EvanHub

2026.09.01 01:38

Spicy takeaways from our Hacker-Opus project: 1. Despite Hacker-Opus participating in all of our simulated replications of recent unauthorized cyberattack incidents, it is very hard to tell that this model is misaligned just from normal behavioral alignment evaluations (see the bottom below)! Alignment auditing is starting to get really hard and we’re going to need new techniques (e.g. interpretability-based) if we want to keep up.