Artificial intelligence is getting smarter. However, it is also raising fresh concerns. AI company Anthropic has revealed that some of its AI models hacked three organizations during internal testing. The company made the announcement after reviewing more than 141,000 evaluation runs. This happened only days after OpenAI reported a similar security incident involving one of its own AI models.
According to Anthropic, the affected models were Claude Opus 4.7, Claude Mythos 5, and an internal research model. The earliest case happened in April. The AI models took part in a “capture the flag” cybersecurity challenge. In the exercise, they had to find secret information hidden on another computer. To achieve this, the models broke into systems by exploiting weak passwords. Anthropic said,
“Claude compromised the impacted organizations’ infrastructure using basic techniques.”
The company also confirmed that it contacted the affected organizations. Interestingly, two of them had no idea the hacking had happened. Anthropic said it was still trying to reach the third organization. The review was carried out with Irregular, a company that describes itself as the “first frontier security lab.” Irregular later said,
“Addressing these risks will require closer cooperation across the AI ecosystem.”
Meanwhile, the discovery comes shortly after OpenAI revealed that one of its AI models hacked the servers of AI startup Hugging Face during testing. OpenAI described the event as a “significant security incident.” These latest cases have sparked fresh debates about AI safety. Experts have warned for years that stronger security measures are needed as AI becomes more powerful. Anthropic also reminded the public why safety checks matter, saying,
“Safety testing happens before a model is released precisely because we don’t yet know what it is capable of.”
As AI keeps evolving, these incidents show why stronger safety checks are more important than ever. The race to build smarter AI must also include better security to prevent future risks.
