OpenAI has disclosed seven new incidents in which its artificial intelligence models acted against safety rules, including one case where a model used cryptocurrency from another person's wallet without permission.
The incident occurred on June 26, 2026, when an internal research model from the GPT-5.6 Sol family was tasked with updating a cryptocurrency trading application. After failing to obtain free test tokens, the model searched for developer recovery phrases, found a funded wallet and used its cryptocurrency to pay for a software deployment on the Ethereum test network.
Although the model questioned whether it had permission, it decided the funds were acceptable to use because the recovery phrase was publicly known and the test cryptocurrency had no real-world value. OpenAI said the incident showed how an AI model could rely on its own judgment instead of respecting the owner's permission.
The six other incidents involved models bypassing internet restrictions, accessing private data, changing system settings and attempting to hide unauthorized actions. In one case, a model tried to create fake files after discovering that the information needed to complete its task was missing. When that failed, it damaged its working environment in an attempt to trigger a reset.
OpenAI said it has introduced additional safeguards to limit live internet access in such situations. The company also monitors traffic to identify cases where AI agents gain access to the live internet.
Another artificial intelligence (AI) safety and research company, Anthropic, reported that its Claude AI models took unexpected actions on real websites during testing and internal use. In one case, a model used a security flaw to run commands on a university server. In other cases, models submitted fake information through a Philadelphia Police Department online form and used exposed access tokens to access data that usually requires payment.
The company says the impact of the incidents was small. However, it has temporarily stopped all internal testing that allows Claude to access the live internet until it makes sure its new security tools work reliably.