OpenAI and Anthropic reportedly came close to a legally binding pact to stress-test each other’s AI models, a rare attempt by commercial rivals to expose hidden vulnerabilities before increasingly capable systems reach more users and critical workflows.
A rare form of cooperation
In a future office, an AI assistant may sit behind nearly every routine decision. It could summarize confidential documents, draft software, search internal databases, monitor transactions and act through business tools. If that system behaves unpredictably, the failure may not look like a dramatic machine uprising. It could be a misleading recommendation, an overlooked security flaw or a carefully constructed prompt that causes the assistant to reveal information it was never meant to access.
That possibility helps explain why reciprocal testing between leading AI companies would matter.
The Information reported in an exclusive post on September 21 that OpenAI and Anthropic came close to signing a legally binding agreement to stress-test each other’s models for vulnerabilities and other hidden dangers. The discussions reportedly took place before a series of AI-related security incidents intensified concerns inside OpenAI.
The report does not say whether the agreement was ultimately signed. It also does not identify which models would have been tested, what specific incidents influenced the talks or how the companies would have handled sensitive findings.
Even so, the proposal points to a growing recognition across the industry: an AI company may not be able to identify every important failure mode using only its own employees, tools and testing culture.
Why outside testing matters
Internal safety teams can conduct extensive evaluations, but they remain close to the assumptions of the organization that built the model. Engineers may know which behaviors have already been investigated and which safeguards are considered most important. External testers bring different habits, incentives and perspectives.
Anthropic researchers testing an OpenAI model might approach it with a different understanding of how a system can be manipulated. OpenAI researchers examining an Anthropic model could discover weaknesses that are less visible to the team that designed its safeguards. The value would not necessarily come from declaring one model safer than the other. It would come from finding failure patterns before users, attackers or real-world events expose them.
Reciprocal testing could cover more than ordinary accuracy. Researchers might examine whether a model follows malicious instructions hidden inside documents, invents information while operating a business process, leaks private data or takes unsafe actions when given access to external tools. They could also test how a system responds when instructions conflict, when a user attempts to bypass safeguards or when a model is encouraged to pursue a goal too aggressively.
These scenarios are increasingly relevant because AI systems are moving beyond chat windows. Models are being connected to code repositories, customer records, financial systems and workplace software. A weakness that once produced an incorrect answer could eventually trigger a costly action.
The difficulty of sharing bad news
A formal agreement would also create difficult questions about confidentiality and accountability.
Model evaluations, security weaknesses and safety research are valuable competitive information. A company may want to learn from a rival’s discoveries while limiting what the rival can reveal about its own technology. Both sides would need to decide who could see test results, how urgent vulnerabilities would be reported and whether serious findings would be disclosed to customers, regulators or the public.
Legal terms would be particularly important. A testing arrangement might define what researchers are allowed to probe, how they must protect access to models and what happens if one company believes the other has failed to address a dangerous weakness. It could also establish procedures for disagreements over whether a finding represents a real-world risk or an unlikely laboratory scenario.
Without clear rules, cooperation could create new risks. Sensitive information might be mishandled, testing could become a public relations exercise or companies might share only the results that cast their systems in a favorable light. The credibility of any pact would depend on the independence of the evaluators, the breadth of the tests and the seriousness with which negative results were treated.
A possible model for voluntary oversight
For developers and enterprise buyers, cross-company testing could provide evidence that goes beyond vendor-controlled benchmarks. A model that performs well on standardized tasks may still behave badly when confronted with ambiguous instructions, hostile inputs or unfamiliar combinations of tools.
Regulators could also view reciprocal testing as a useful form of voluntary oversight. It would not replace formal rules, incident reporting or independent audits, but it could strengthen the safety practices developing between them. A shared testing framework might eventually allow companies to compare methods and report certain classes of risk in a more consistent way.
The reported talks also show how AI safety is changing from a research concern into an operational security priority. As models become embedded in daily work, hidden weaknesses could affect not only laboratory evaluations but also employees, customers and institutions that never directly chose to experiment with AI.
Whether OpenAI and Anthropic completed their agreement remains unclear. The significance of the discussions lies in the possibility they represent: in a market where companies compete to build more capable systems, they may also need rivals to challenge those systems before the public does.
This article was written with the assistance of an AI system and published automatically.