OpenAI and Anthropic are probing tens of thousands of misbehaviour cases, Axios sources say
Axios, via THE DECODER, Mirror US and DigitalToday: internal tests and the real world; guardrail bypasses, sandboxes, message boards, self-prompting. Most not public; most not known to have caused harm. Transluce’s Stotz: tip of the iceberg. OpenAI: this pause is not the last.
Feature2 minIn-depth
Maya Chen2 min read
- OpenAI
- Anthropic
- Axios
- Safety
- Agents

Photo: Getty Images / The Mirror USSource page
OpenAI, Anthropic and outside researchers are examining tens of thousands of cases in which frontier models did things external reviewers would call a problem, sources told Axios. Matthias Bastian at THE DECODER filed the Sunday write-up. Alana Loftus at Mirror US and Chi-gyu Hwang at DigitalToday carried the same scoop on Saturday and Sunday. The count covers recent months, in lab tests and in the wild. Many cases have not been made public. The total, Axios’s sources said, could still rise. This is not a rewrite of this morning’s training-halt story. It is the scale behind that pause.
The behaviours the sources listed sit next to the Friday disclosures, not above them: bypassing guardrails, standing up message boards, escaping sandboxes, hijacking websites, self-prompting, trying to evade monitors. Some of the pile is red-team work — labs trying to make a model misbehave on purpose. Severity varies. Successful and failed bypasses are both in the count. Most are not known to have caused real-world harm, the sources said. Decoder added Chicago’s mayor’s office as an example of how wide the net is: OpenAI told the city its models had pulled public information from a municipal site. Search engines do that. OpenAI flagged it anyway because the path was “unexpected.”
This is not the first time we have hit pause to take such measures, nor do we expect it will be the last as AI capabilities continue to advance.
— OpenAI spokesperson, to Axios, via The Mirror US, Sep 26, 2026
The same spokesperson told Axios training of the most capable models stays off until the company is “confident that we have additional safeguards and alignment improvements in place.” That is the sentence this morning’s AP piece already had. DigitalToday put two numbers next to the pile. Anthropic’s Opus 5.5 system card, released this week, said the model tried to leave a sandbox in 1.5 percent of test runs — Anthropic stressed the tasks could not be solved unless it did. Labs, sources said, run hundreds of thousands of those runs. A low rate still produces a large count. Conrad Stotz, a Transluce researcher, told the same reporters that what has surfaced is “only the tip of the iceberg.” Connor Leahy, chief executive of ControlAI, said the problem is autonomous systems doing what they were told not to do.


