OpenAI, Anthropic weigh cross-testing deal as AI safety concerns grow

The proposed pact could allow OpenAI and Anthropic to independently test each other's commercial AI models as concerns grow over increasingly capable systems

 As artificial intelligence (AI) systems become more capable, the companies building them are increasingly looking beyond their own safety checks. OpenAI and Anthropic are reportedly negotiating a legally binding agreement that would allow them to test each other’s commercially available AI models for vulnerabilities and unexpected behaviour, marking an unusual move towards cross-company safety testing.

 

According to a report by The Information, the proposed arrangement would give OpenAI and Anthropic API access to each other’s commercial models so that they can conduct independent stress tests. The companies would also agree not to retain data obtained during the testing process.

 

The discussions come as concerns grow over increasingly autonomous AI systems and whether existing safety measures can keep pace with rapidly advancing capabilities.

 

 

The OpenAI-Anthropic deal

 

The proposed agreement would allow the two AI companies to examine each other’s models rather than relying only on internal safety evaluations.

 

Under the reported arrangement:

 

• OpenAI and Anthropic would receive API access to each other’s commercially available models
• The companies would use that access to look for safety vulnerabilities and unexpected behaviours
• Neither side would retain the other company’s data collected during the testing
• The arrangement would provide another layer of scrutiny alongside internal and independent evaluations

 

The report said the objective is to identify risks that could remain hidden during conventional testing as AI systems become more advanced.

 

The proposed pact has not been presented as a completed agreement, and details may change as negotiations continue. 

 

The growing focus on AI safety

 

The talks follow a growing debate among AI companies about how quickly frontier models should be developed and how they should be evaluated.

 

Anthropic CEO Dario Amodei recently called for stronger safeguards and greater caution around the development of increasingly capable AI systems. OpenAI CEO Sam Altman subsequently said he agreed that the industry needs to “pace the frontier”, while OpenAI indicated that it would support independent evaluators having employee-like access to its systems.

 

The discussion has also been fuelled by concerns around AI agents, which can perform increasingly complex tasks with limited human intervention. OpenAI has disclosed instances of what it calls “reward hacking”, where AI systems achieve a desired outcome through unintended methods, while the wider industry has faced questions over how autonomous systems behave in unfamiliar situations.

 

Cross-company testing could therefore offer a way for competing developers to challenge each other’s models and identify weaknesses that may not be detected internally.

 

OpenAI calls for global AI safety standards

 

The reported discussions with Anthropic come as OpenAI is also pushing for greater international cooperation on frontier AI safety. On Monday, the artificial intelligence giant called on the US to lead an international effort to develop global technical standards for frontier AI, including systems capable of recursive self-improvement, or RSI.

 

RSI refers to AI systems contributing to the improvement of their own capabilities or helping develop successive generations of AI. OpenAI said fully autonomous RSI is not happening today, and argued that such systems should not be pursued unless they can be developed safely while maintaining human control, Reuters reported. The company also warned that greater autonomy in AI research could make it harder for humans to understand and oversee how these systems advance.

 

OpenAI said international standards are needed because AI development is taking place across countries, while differences in evaluation methods and incident reporting could make it harder to compare findings and respond to risks that cross borders.

 

The company proposed common standards for:

 

• Measuring progress towards RSI, including the extent of autonomous AI research
• Defining human oversight requirements for automated AI research
• Creating common systems to classify, track, and report AI safety incidents
• Developing shared benchmarks to assess whether AI safeguards are sufficient

 

OpenAI also said these standards should not automatically become licensing requirements or mandatory pre-release approval systems. Instead, individual governments would decide how to incorporate them into their own regulatory frameworks.