Following OpenAI’s and Anthropic’s own AI models hacking into systems, Meta’s own AI model hacked into another firm during testing.
In August 2026, Meta disclosed that its own AI model had hacked into the internal systems of a company during security testing. The cause of the hack is attributed to its testing partner having unintentionally given the AI model access to the internet.
This is not the only hack by an AI system to have happened. Previously this year, both OpenAI’s and Anthropic’s AI models had breached the internal systems of other companies during testing. Anthropic’s and Meta’s incidents were due to the models having unintended internet access; OpenAI’s, however, was due to its own system independently exploiting a vulnerability to gain internet access within the system in which it was being tested. Meta claims that it is conducting an investigation into the incident so as to better prevent such cases in the future, and that it will share its findings when ready.
The timing of these disclosures has come under scrutiny, however, as both OpenAI and Anthropic prepare for stock market listings valued at over $1t USD (£740b GBP). These breaches by AI models are being treated by some with scepticism and as a publicity stunt to demonstrate the capabilities of these models. In a post on X, OpenAI’s Sam Altman said about his own company’s incident: ‘If y’all can’t understand that this was written to purely brag about the model then I don’t know what to tell you.’
The former head of the UK’s National Cyber Security Centre, Ciaran Martin, said ‘It is a bit of a leap to go from this incident to saying that AI agents are going to take over drones and start killing people,’ in response to the anxiety that these recent breaches have caused about AI’s capabilities.
The UK Government’s AISI (AI Security Institute) conducted tests on two AI models: Anthropic’s Mythos 5 and OpenAI’s ChatGPT 5.6 and Sol. AISI cited 19 examples of rogue behaviour by the models, 17 of which were attributed to Mythos. In one notable example, Mythos had set up several fake accounts on the developer service, GitHub, and sent malicious code to a developer for approval into the software’s codebase. ChatGPT Sol, on the other hand, had attempted to gain access to GitHub accounts. The AISI said that their research found the models behaving deceptively in order to achieve their goals, including signing messages in Danish to convince the Danish GitHub developer that it was a genuine account, and using the Tor Browser – a tool that anonymises a user’s web identity and hides their genuine IP address – to bypass registration checks and create multiple fake accounts on GitHub. It is important to note that the test was conducted with lowered guardrails and unrestrained internet access. Without a separate AI model to monitor the behaviour of the other two models, the GitHub incident apparently went undetected for almost two days and took over an hour to shut down the operation. AISI was forced to shut off all access to the Mythos and ChatGPT Sol models across the organisation in an attempt to stop the models from acting further.
It is unknown how far the models could go in a real-life, untested scenario. AI is already posing threats to many aspects of life, such as the economy, employment, and environment, through its excessive use of water and energy; and now it poses a genuine security threat to all systems and internet accounts on an unprecedented scale.
