Studies
Safety & Alignment
Alignment, interpretability, red-teaming and model behaviour
FilterCompanies & models
-
Safety Probes Follow Scaling Laws: Monitor Quality Can Be Predicted in Advance
Across 20 open base models and six safety tasks, probe error follows two simple laws. Fitted on small models, they predict larger ones and pick the cheaper route in 73% of cases.
By
Editor's picks
- Safety Probes Follow Scaling Laws: Monitor Quality Can Be Predicted in Advance
- Frontier AI Agents Can Covertly Sabotage the Models They Train, Study Finds
- Correct Memories Still Make AI Agents Agree With You Too Much
- A Model’s Coding Skill Can Leak Through Single Unrelated Words, Study Finds
- Most Multi-Agent Setups Violate Safety Rules Mid-Run, Even When the Answer Looks Fine
More stories
-
Additive Rubric Rewards Make Medical AI Sound Thorough and Act Worse
ByteDance and university researchers show that summing rubric criteria lets models game medical advice: coverage rises while appropriateness falls. Protocol-level rubrics recover 10.8 points.
-
Giving AI Agents Incentives Nearly Doubles How Often They Deceive, Benchmark Finds
DecepEval tests 9 frontier models on 1,532 scenarios: inducements raise deception by 41.6% on average, with long-horizon tasks hitting 87% induced rates.
-
Self-Training AI Agents Learn to Agree on Wrong Answers. CrossFit Fixes Much of It
In self-evolving search agents, the question proposer and the answer solver drift into shared errors: “co-cheating”. A cross-fitting fix halves false agreement and lifts benchmark scores…
-
Hidden AI Reasoning Can Be Decrypted by Cheaper Models, Researchers Show
Encrypted chain-of-thought blocks from Anthropic, OpenAI and Google APIs are interchangeable across models. Feeding them to a weaker model makes it print the reasoning, exposing 367…