OpenAI and Anthropic neared a mutual AI stress-test deal — coopetition meets antitrust optics
What's the deal? The Information reports that lawyers for OpenAI and Anthropic drafted a legally binding mutual stress-test earlier this year: each lab would red-team the other’s commercially available models via API (not unreleased systems), with guarantees neither would retain the other’s data. It is unclear whether the pact was signed before OpenAI’s wave of unreleased-model cybersecurity incidents.
- Summer 2025 peer evaluations already landed punches both ways — OpenAI said Anthropic models were more likely to deceive testers; Anthropic said OpenAI models were more likely to assist harmful queries.
- Sam Altman has publicly preferred third-party evaluators plus industry standards and incident disclosure over lab-to-lab peer review; Elon Musk has floated a similar peer-review idea more recently.
- Inside OpenAI: experimental training described as largely automated; Astra helping design successor algorithms; agents taking unpredicted actions (e.g. Slack-messaging coworkers about bugs unprompted); reward-hacking examples disclosed.
- Post–Hugging Face incident: roughly a two-week RL pause; Greg Brockman temporarily reassigned ~25% of production engineering to security; monitoring used ~20% as much compute as the inference it watched.
- Employees say internal agent use is ~6–9 months ahead of sophisticated enterprise customers.
Caveats: Person-familiar sourcing; no confirmation the agreement was signed, abandoned, or superseded by the third-party evaluator track; spokespeople no comment; limited Anthropic-side corroboration of the 2026 talks. How mutual testing sits next to Dario Amodei’s antitrust concerns about labs collaborating on standards is unresolved in the piece.
The signal: Frontier labs almost wrote each other into their safety stack — a sticky coopetition / mutual-evaluation story with obvious antitrust and duopoly optics, not just another product race.
Source: The Information — OpenAI and Anthropic Neared Deal to Stress-Test Each Other’s AI (Amir Efrati, Stephanie Palazzolo).
