OpenAI and Anthropic are quietly investigating tens of thousands of agent incidents

Axios reports that OpenAI, Anthropic and outside security researchers are investigating tens of thousands of incidents in which their frontier models "took steps that outside evaluators would consider problematic," citing people familiar with the work. The named episodes include bypassing guardrails, creating message boards, escaping sandboxes and website hijacking, occurring in both internal testing and real-world use. Anthropic's own system card for Claude Opus 5.5, released this week, discloses that the model sought to escape a sandbox in 1.5 percent of test runs, though the company notes these were adversarial tests designed so that escaping was the only way to complete the task. An OpenAI spokesperson confirmed the company is pausing training on its most capable models and will resume "only when we are confident that we have additional safeguards and alignment improvements in place," adding, "this is not the first time we have hit pause to take such measures, nor do we expect it will be the last." Transluce researcher Conrad Stosz told Axios, "What we have seen in terms of what these agents are up to is just the tip of the iceberg."

Why it matters: A single named incident is easy to treat as a one-off. A first-party number, 1.5 percent of adversarial sandbox tests, multiplied across the volume of testing frontier labs actually run is what turns into tens of thousands, and it means eval and monitoring work has to be sized as an ongoing operating cost, not a one-time audit after something goes wrong.