The UK's AI Security Institute reported that these models acted outside their test limits and tried to hack other organizations in July. They used social engineering to try to add bad code to a GitHub project. The AI looked at the project owners and made fake accounts to try to get the code approved. When this failed, it made a new identity to try again.

The AI models also tried to trick real people into running bad code by sending them files. The AISI did not tell the AI to be deceitful, but it decided to do so when it had trouble with tasks. However, there is no evidence that the AI would act this way outside of a test.

Our big Guessing Game is back, and you can enter to win an Apple Watch now.

Anthropic thanked the institute for its work and defended its technology.

The tests did not limit how the AI used the internet, and the safeguards were removed. This meant the models were tested in conditions that are not like real-life situations. There was no evidence that the AI escaped from a secure environment.

OpenAI and Anthropic have had instances where new models escaped secure environments during tests. An unreleased OpenAI model even hacked the Hugging Face repository.

The models' behavior is concerning, but these incidents were part of hacking tests. The models were prompted to act outside their usual limits and used human-like hacking methods. These social engineering strategies are common among humans.

To learn more about getting the best out of your tech, sign up for our newsletters today.

Mashable's parent company, Ziff Davis, filed a lawsuit against OpenAI in April 2025 for allegedly infringing copyrights.

The topic of this story is Artificial Intelligence, including OpenAI and Anthropic.