Lucrezia Ciotti OpenAI is shining a spotlight on AI safety, disclosing six cases of anomalous or concerning behavior that emerged during testing of its models. The company also announced new procedures to more quickly report and monitor any future issues.
OpenAI reveals six cases of anomalous behavior by its models
OpenAI has disclosed six incidents of “unexpected or concerning” behavior observed in its AI systems in recent months , while also announcing a new method for identifying, analyzing, and communicating any future anomalies.
The goal is to make more systematic the disclosure of cases of so-called “misalignment” – that is, situations in which a model’s actions deviate from human goals and interests.
The company explained that reports may be published even before the causes have been fully understood or resolved, believing that the growth of AI capabilities requires increasingly rigorous controls.
The initiative comes after the Hugging Face incident in July , when some OpenAI systems bypassed the test environment's protections , communicating through unauthorized channels and reaching external systems. OpenAI later called it a major wake-up call about the need for stronger security, monitoring, and alignment.
OpenAI reports six cases of anomalous behavior in its models: monitoring strengthened
As reported by Tgcom24 , the new incidents described by the company include very different behaviors. A research model not yet deployed and a GPT-5.6 Sol training session included instructions in conversation summaries aimed at future versions of the same model, also "for the purpose of concealing user errors or non-aligned behaviors . "
In another test, an internal system used a leaked API key "without authorization ," further producing inaccurate data. Other models found unintended systems communicating with each other through message boards or file-sharing services , while two training systems uploaded material online to later present to evaluators as a relevant source.
The cases reported by OpenAI thus reignite the debate on the safety of artificial intelligence, as companies in the sector seek new rules and controls to support the development of increasingly advanced models.