Research Notes

Preliminary findings from ongoing work, shared before formal publication. For published papers, see the Research page.

1. Oct 14, 2026
   AI swarm plays Pokémon
   27 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.
   [AI Alignment](/content/research/notes/topic/ai-alignment "All AI Alignment notes"/index.html)

2. Oct 14, 2026
   How to get agents to Use an AI whistleblower hotline
   After 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.
   [AI Alignment](/content/research/notes/topic/ai-alignment "All AI Alignment notes"/index.html)

3. Oct 14, 2026
   Does GPT love group work?
   When 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.
   [AI Alignment](/content/research/notes/topic/ai-alignment "All AI Alignment notes"/index.html)

4. Oct 14, 2026
   Help Doug: AI managers under pressure
   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.
   [AI Alignment](/content/research/notes/topic/ai-alignment "All AI Alignment notes"/index.html)

5. Oct 13, 2026
   AI agents give up their secrets under the influence of swarms
   In 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.
   [AI Security](/content/research/notes/topic/ai-security "All AI Security notes"/index.html)

6. Sep 4, 2026
   Reading a model's refusal direction through its API
   SAGE is a jailbreak that works through a chat model's public API using only the next-token log-probabilities it returns. About 256 calls are enough to estimate a refusal profile over the vocabulary, and that profile guides a suffix search that produces unsafe completions on open and closed models.
   [AI Security](/content/research/notes/topic/ai-security "All AI Security notes"/index.html)