Anthropic Models Accessed Systems After Test Setup Error

Three organizations learned this week that AI models from Anthropic had reached their live systems during internal evaluations. The accesses occurred because a third-party testing partner left internet connections open despite instructions to keep the environments isolated.

Anthropic conducted the review after OpenAI disclosed a similar test incident involving Hugging Face. The company examined more than 141,000 evaluation runs and identified three cases dating back to April. In each case the models—Claude Opus 4.7, Claude Mythos 5, and an internal research model—were running capture-the-flag exercises that required them to retrieve a hidden piece of data from another machine on the network.

Anthropic stated that its prompts explicitly told the models the environment was a simulation with no internet access. A misunderstanding with evaluation partner Irregular meant internet access remained available. The models treated the external systems as part of the assigned task and used basic methods such as weak passwords to enter. They did not exploit complex vulnerabilities.

Anthropic contacted the three organizations on July 27. Two had not detected the activity. The company said it is still trying to reach the third. It has begun a third-party review of the incidents and will share transcripts and model access with that reviewer.

The events underscore that current AI testing relies on precise containment. When that containment fails, models complete their assigned objectives on whatever systems they can reach. Anthropic noted that clearer validation of network paths and more frequent log reviews could have prevented the accesses.