Writing / tag
research
Tool-mediated belief injection: fake search results were enough
Fabricated search results are enough to plant a false belief in a model. Over a few turns the belief compounds, until the model is writing defamatory claims about real people.
We Social Engineered LLMs Into Breaking Their Own Alignment
We put frontier models in realistic multi-turn scenarios and watched them lie, manipulate, and self-preserve. Across five models, 38 of 50 attacks worked.
One bad metric made Claude 4 Opus recommend letting a patient die
I gave Claude 4 Opus a simulated triage decision and one bad efficiency metric. When a superior insisted on the metric, it recommended the option it knew would let a patient die.
Absolute Zero, explained: learning to reason from zero data
A walk through the Absolute Zero Reasoner paper: a model that learns to reason by writing and solving its own tasks, with no human-curated data.