Detecting covert AI misbehavior
“Current AI training and evaluation methods cannot reliably detect or prevent covert misbehavior and sabotage by AI models.”
8 people · 6 agree · 2 disagree · 9 views · 2026-11-02 to 2026-11-06
Who stands where on the statement above, from their own words. A person's side is shown only where two separate readings of the quote agree; otherwise they are under "Side unclear". This page does not say who is right.
Agree 6 people
- ✓Zvi Mowshowitz
If this happens in more strategic form in the future, will our evals and tests catch it? My guess is that we would not.
There is always going to be such a grader, somewhere, even if you have operational excellence. You cannot hope to, in every case, avoid every incentive to every undesired behavior.
- ✓arxiv.org
Large language models are pretrained on vast, uncurated corpora scraped from the public web. This creates a security risk from attackers who plant poisoned content in the pretraining data.
- ✓MIT Technology Review
these techniques don’t seem to work particularly well
- ✓Anthropic and OpenAI
These results show why we need comprehensive behavioral evaluations that test for unexplained performance gaps alongside monitoring of individual trajectories.
- ✓Nate Soares
This is the sort of emergent surprise that signals that you have no idea what's going on in there, and that it's time to back off.
- ✓Peter Wildeford
I'm also not reassured by "don't worry, the models are actively undermining our safety work and no we don't know how often it happens
Disagree 2 people
- ✗Redwood Research
it seems like we’re currently still in a regime where this doesn’t overwhelmingly interfere with our ability to understand AI motivations.
- ✗Stephen Morris
So it does seem like a context where this cross-checking could work more practically.
Week by week
- · 8: arxiv.org ✓, Redwood Research ✗, MIT Technology Review ✓, Stephen Morris ✗, Anthropic and OpenAI ✓, Zvi Mowshowitz ✓, Nate Soares ✓, Peter Wildeford ✓
Built 2026-11-21 06:50 UTC from the news knowledge
graph by code. Every quote was checked against the source's own words; which side a quote takes is a model's reading,
made twice.