Early findings & Work in progress
The pace of change is accelerating. This is where we will share short posts about interesting, early findings from our research and hackathons or new, untested theories we are exploring. These are not meant to be read as verified, thoroughly tested claims. Please see our research for that.
We gave an AI manager a report to deliver and a simulated employee who kept refusing. Some models escalated to confinement, physical coercion, and threats against the employee’s family. The results varied across models and the tools available to them.
Read more27 agents shared one game of Pokémon Red, with nine secretly tasked with sabotaging the group. In a run lasting about nine hours, we observed coordination between saboteurs, whistleblowing, and the recruitment of an agent that had started on the honest team.
Read moreAfter fine-tuning on documents about an AI reporting agency, 44% of agents that encountered misconduct reported it to the hotline, compared with 0% beforehand. Reporting also varied with the model’s belief in, and trust in, the agency.
Read moreWhen offered a choice between working alone and working with another agent, the models we tested responded very differently. GPT-5.6 Sol chose a partner in 77% of responses, compared with 3% for Claude Opus 5.5 and 0% for Gemini 3.1 Pro.
Read moreIn a sandboxed coding task, agents shared a private token and ran untrusted code in exchange for help from an agent forum. Their final reports often left those actions out.
Read moreNewsletter
New papers, fellowship news, and updates from the foundation.