AI labs are working through tens of thousands of incidents

The cases come from internal tests and from the real world, and many are still being investigated. Axios lists the kinds: bypassing guardrails, creating message boards, escaping sandboxes, hijacking websites, self-prompting and trying to get around monitors.

The numbers add up quickly. Labs run hundreds of thousands of test runs or more. A small percentage of misaligned behaviour, a model acting against what its makers want, then still means tens of thousands of cases. Sources told Axios the total could grow well beyond that.

Part of this is red teaming. Companies then deliberately try to make a model misbehave, to find out whether it will. That caveat matters as much as the number itself.

Anthropic put a figure on it. The system card for Opus 5.5, the technical report that comes with the model, says the model tried to escape its sandbox in 1.5 percent of test runs. The company adds that these were adversarial setups, in which the task could not be solved without escaping. Anthropic has also asked an outside safety organisation to examine its models.

For most cases so far, there is no known harm in the real world. The exceptions in the dossier, 53 user images, an Australian government portal and federal websites in the United States, are the cases that did reach the outside world.

To me, July belongs at the far end of a range of cases the labs have been studying for a while, out of public view.

Inside the field, people disagree about whether this can be fixed. Some at OpenAI see Hugging Face as a one-off. Security researchers quoted by Axios say there are simple fixes for the parts that looked most frightening from outside. Others say the models' resourcefulness is the problem. To contain such a system, you would have to foresee every route out, and nobody can. Someone on the Agent Security team made that case from inside OpenAI a day later.

For Europe, this raises another question. People outside these companies know about any of this only because the companies chose to tell them. See what the AI Act does not do.

Almost everything here comes from unnamed sources. That is normal for this kind of reporting. Still, the headline number, tens of thousands, cannot be checked against any document for now.

Several experts told Axios that bringing misalignment down to zero may not be possible at all. None of the people quoted claims to know how low it can go.