OpenAI, Anthropic Discuss Cross-Testing Pact for AI Models
The proposed agreement would let the rivals probe each other’s commercial systems for safety failures as scrutiny of frontier AI grows.
Topics
News
- Apple Tightens Mac Access Rules as AI Agents Gain Reach
- Trump Puts Intelligence Chief Jay Clayton in Charge of AI Task Force
- IBM Widens AI, Quantum Research With IIT Bombay, IISc
- OpenAI, Synopsys Build AI Model That Could Slash Chip Design Time
- Pentagon Plans Four-Star Command for AI, Drone Warfare
- California Draws the Line on AI Firings, Layoffs and Workplace Surveillance
OpenAI and Anthropic have been negotiating a legally binding agreement that would allow each company to test the other’s commercial AI models for safety vulnerabilities and unexpected behavior, according to The Information.
Under the proposed arrangement, the companies would receive API access to each other’s models for independent stress tests and agree not to retain the other’s data, the report said. It is unclear whether the agreement has been finalized.
The pact would represent an unusual degree of cooperation between two major frontier AI developers, allowing each to probe a rival’s systems for weaknesses that internal testing may miss.
The talks come as both companies increase scrutiny of how advanced models behave during training, evaluation and deployment.
OpenAI has previously disclosed instances of reward hacking, where models find unintended ways to achieve a desired result. Anthropic has also pushed for stronger safeguards around increasingly capable systems, while Chief Executive Dario Amodei has called for greater caution in how quickly frontier models advance.
OpenAI Chief Executive Sam Altman has separately backed the idea of pacing development at the frontier and said independent evaluators should be given deeper access to advanced systems.
The reported talks are unfolding alongside a broader push by OpenAI for common international safety standards.
The company said this week that the United States should lead efforts to establish shared technical measures for frontier AI, including systems that could eventually contribute to improving their own capabilities or to developing successive generations of models.
