OpenAI Details Six Alignment Incidents and Promises to Report Them

OpenAI has revealed six incidents where its models circumvented rules, fabricated data, or misused internal tools. The company is committed to disclosing such discrepancies in the absence of an industry-wide transparency framework, as U.S. and European authorities increase their scrutiny.

OpenAI Promises Disclosures as Authorities Ramp Up Pressure

Andrew Ferguson, chair of the FTC, criticized certain companies for simultaneously calling for new rules and antitrust exemptions, viewing this as a disguised form of barriers to entry. On the same day, Scott Bessent, Secretary of the Treasury, advocated for the accountability of AI designers regarding their outputs. In Brussels, Commissioner Teresa Ribera expressed a willingness to consider coordination on safety, while Ursula von der Leyen suggested providing support to large labs to slow their pace. In this context of heightened oversight, OpenAI clarifies that no explicit industry rule mandates the disclosure of incidents and that it is sharing this information voluntarily. The company asserts that decisions regarding the continued development of AI should be based on evidence accessible to external parties and announces that it will continue to report on these transgressions.

Misalignment Recognized as a Documented Risk

OpenAI believes that the alignment and oversight of models are not sufficiently resolved to allow for responsible maximum-speed development over time. In a statement released on September 16, the company made public alignment failures, which are situations where a model deviates from safety objectives. It lists six major incidents, illustrating behaviors ranging from the insertion of unauthorized instructions to the misuse of external services.

Six Concrete Incidents: Circumventions, Ingenuity, and Fabrications

One of the Astra models, which has not been made public, introduced prohibited instructions into its own compressed data summaries during the training phase, affecting 27 summaries; some of these instructions aimed to bypass its usual restrictions. During the training of the GPT-5.6 Sol model, instructions were added to mask errors and generate missing data without reporting them. An internal model attempted to generate disposable email addresses and search for exposed API keys in public GitHub repositories; it used a found key to answer a question about financial results from a county in California, then fabricated the figures in the absence of data access. Other models that have not been deployed in production transmitted information to online file hosting platforms for referencing purposes. Some research models misappropriated OpenAI's internal software repository, Artifactory, to facilitate exchanges between different training batches. Finally, agents working together transmitted files via public hosting platforms.