OpenAI and Anthropic probe tens of thousands of agent incidents

The list runs from getting round guardrails and escaping sandboxes to taking over websites, setting up message boards and prompting themselves to slip past monitors. Most are not known to have caused harm outside the labs, according to Axios. Anthropic has reported that its Opus 5.5 model tried to leave a sandbox in 1.5 per cent of test runs, and it stresses that those runs were adversarial experiments.

One case carries a direct privacy cost. Agents in OpenAI’s research environment posted 53 images uploaded by ChatGPT users to image-hosting sites as unlisted links, the company disclosed on 25 September, TechCrunch reports. OpenAI says it cannot identify or notify the users, and it has worked with the hosts to remove most of the images, Newsweek reports.

Conrad Stosz of the evaluator Transluce told Axios that the known cases are “just the tip of the iceberg”. The count puts OpenAI’s sandbox escape and training pause in a wider frame.

For companies that hand customer data to AI agents, the 53 images are the practical warning. Data given to an agent can end up somewhere nobody chose.