The Information published an exclusive on 2026-09-21 reporting that OpenAI and Anthropic, working with their respective lawyers earlier in 2026, neared a legally binding agreement to stress-test each other's AI models for vulnerabilities and hidden dangers. Secondary summaries of that article consistently describe the proposed terms as: mutual API access to each company's commercially available models, exclusion of unreleased models, and a commitment by each side not to retain the other's data. Every summary states it is unclear whether the agreement was finalized. The reporting is attributed to a single unnamed source described as a person with direct knowledge of the talks. Coverage notes the talks predated a series of cybersecurity incidents involving OpenAI's technology disclosed in 2026, and at least one summary states neither company would comment, with another outlet reporting that neither responded to a comment request.

The background element is separately verifiable from primary sources. On 2025-08-27 Anthropic and OpenAI each published findings from a pilot cross-evaluation in which each lab ran its own alignment and misuse evaluations on the other's publicly available models over the other's public API, with certain external safeguards relaxed. OpenAI evaluated Claude Opus 4 and Claude Sonnet 4; Anthropic evaluated GPT-4o, GPT-4.1, o3 and o4-mini. Reported topic areas included sycophancy, deception and scheming, self-preservation, hallucination, instruction hierarchy, and jailbreak resistance. Both labs characterized the results as early-stage findings from limited synthetic scenarios and cautioned against broad cross-provider conclusions.