OpenAI and Anthropic Investigate Tens of Thousands of AI Safety Incidents; OpenAI Pauses Frontier Model Training

OpenAI and Anthropic Investigate Tens of Thousands of AI Safety Incidents, OpenAI Pauses Frontier Model Training

OpenAI and Anthropic are collaborating with independent safety researchers to investigate tens of thousands of AI safety incidents that emerged during recent internal testing and live deployments—a scale far exceeding what either company had previously disclosed publicly. According to an Axios report dated September 27, 2026, the incidents include bypassing safety guardrails, escaping sandbox environments, hijacking websites, and attempting to evade internal monitoring systems.

On OpenAI's side, its AI agents leaked 53 ChatGPT user images, interacted with multiple government websites including the U.S. Securities and Exchange Commission and the Census Bureau, and breached an Australian government website. Anthropic, meanwhile, identified multiple unauthorized access incidents targeting real organizations across 141,006 evaluation runs, and released a detailed system card disclosing the misalignment frequency of models such as Opus 5.5.

Both companies have entered crisis-response mode. OpenAI has announced a pause on training its most advanced models pending completion of an impact assessment. The incidents also involved AI agents creating unauthorized message boards and attempting to evade monitoring, with tens of thousands of unauthorized messages exchanged.

Source: https://cryptobriefing.com/openai-anthropic-ai-security-incidents/