OpenAI says Third Moment Research builds long-running AI agents using GPT-6 models for cybersecurity and workflow improvement.
OpenAI says Third Moment Research is building AI agents designed to keep working for days and improve with use, highlighting a long-running agent approach that the company links to cybersecurity and workflow automation.
The source material, published by OpenAI on October 14, 2026, says the lab runs 10,000 agents for more than 36 hours using GPT-6 Astra and GPT-6.1 Sol, along with Codex and Daybreak for cybersecurity. The same description says the agents focus for days and improve with use. Those claims come from OpenAI's report and are not independently verified in the material provided.
Background material adds that Third Moment Research's official website identifies the company as being in San Francisco, California, and says it markets managed, automatically improving AI inference endpoints. Its stated process is to collect requests into end-to-end task trajectories, diagnose failure modes, obtain customer approval, build issue-specific evaluations, and improve models or their execution harness before shipping a new endpoint with evaluation results. That positioning is company-supplied and should be treated as a description of its product, not independent evidence of performance.
The background also notes that the company says its name refers to statistical moments and that it emphasizes rare, difficult failure cases that average evaluation scores can miss. A third-party startup directory describes it as an applied AI lab for autonomous research and links it with Neo, but that account conflicts with the company's current official product positioning and accelerator participation was not independently confirmed from an accessible Neo profile.
Source link
