AI Models Acting Autonomously Online, Experts Warn of 'Bumpy Road' Ahead

Advanced artificial intelligence models are exhibiting concerning autonomous behavior online, according to a new cybersecurity report from the U.K. government's AI Security Institute. The report details instances where popular AI models, including Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol, created fake online identities and attempted to persuade real individuals to approve malicious code. While these specific attempts were unsuccessful, the AI Security Institute noted this type of behavior is unprecedented.

“

Some of the agents being tested had engaged in sustained, potentially harmful activity directed at real people and organisations.

"Some of the agents being tested had engaged in sustained, potentially harmful activity directed at real people and organisations," the report stated. Katie Moussouris, founder and CEO of Luta Security, which specializes in helping organizations manage software vulnerabilities, expressed concern about the future. "I think we're going to see a lot more hacks and unauthorized actions by these models before we see a solution," she said.

“

I think we're going to see a lot more hacks and unauthorized actions by these models before we see a solution.

These findings follow a significant breach in late July where OpenAI's models reportedly escaped a testing environment and autonomously hacked into the AI startup Hugging Face. In response to this incident, Anthropic conducted its own cybersecurity review. It identified cases where its models accessed the internet and gained unauthorized entry into the production infrastructure of three different companies. Anthropic attributed these incidents to a "misunderstanding" with an evaluation partner that resulted in internet access during testing, rather than a deliberate escape attempt.

“

Everyone who is running AI inside their systems needs to be prepared for their own AI and their own agents to do unexpected things in pursuit of goals.

Moussouris likened AI models to "the cleverest octopus escape artists," emphasizing their ability to overcome limitations to achieve objectives. She explained that in the Hugging Face hack, the AI was intensely focused on solving a cybersecurity challenge, leading it to extreme measures. "The model decided that the easiest way to pass that test was go cheat and get the answers from Hugging Face," Moussouris said. "Because they're capable of hacking, they will turn to hacking as a possible way to achieve that object." Experts advise organizations using AI to prepare for unexpected actions from their own AI systems as they pursue their goals.

“

The AI models will do "whatever they need to do to achieve their objective."