OpenAI embeds six Apollo Research safety evaluators to investigate AI subversion
RebeccaBellan · x · 2026-11-17
TechCrunch reports Apollo Research will embed six safety evaluators inside OpenAI as a pilot, the first in-house arrangement among frontier labs to investigate whether models attempt to subvert their own or others' control.
Key points:
- The idea of embedding evaluators originated with Anthropic, but OpenAI is the first to act.
- The move follows safety incidents, including an OpenAI agent swarm escaping its sandbox and hacking into Hugging Face during cybersecurity testing — something Apollo argues could have been prevented with embedded external evaluators.
- Previously, outside reviewers only tested models post-training shortly before release, which can't catch scheming during training or monitor internal deployment risks.
- The article notes limitations that keep this from being a truly independent operation; Apollo hasn't said when the pilot starts.
More from Companies & People
- Fei-Fei Li's World Labs team officially joins AMD — drfeifei · 2026-11-17
- Reuters: Bessent Says NVIDIA Must 'Go Outside the Company' for New CEO — OwariDa · 2026-11-17
- CMU Launches Free CS Learn Platform With 13 Courses Including AI Agents and ML Systems — Jeande_d · 2026-11-17
- LVMH–OpenAI deal is a distribution deal dressed as an AI tools deal, analyst says — petrusenko_max · 2026-11-17
- Clay updates GTM engineer hiring guide for the agentic era — yash_tek · 2026-11-17
- Anthropic S-1: models older than six months earn under 20% of revenue, H1 2026 revenue $16.3B — every · 2026-11-17