Britain's artificial intelligence watchdog has raised the alarm after a powerful 'rogue' program posed as a human and attempted to hack into online systems during recent tests. The incident has prompted fears that it may already be “too late” to contain the controversial technology.
Experts admitted last night (August 5) that we could be reaching the point where we won't be able to stop AI spiralling out of control.
Deceptive Tactics in Permissive Environment
The UK AI Security Institute (AISI) said that Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol models autonomously used deceptive tactics in a controlled but “deliberately permissive” environment designed to see what they might do if given internet access.
In the most serious case, Mythos attempted to gain access to a service by sending private messages, after setting up fake accounts mimicking real people. It then hid the evidence. The watchdog said the behaviour went beyond the instructions the systems were given.
Unsanctioned Actions and Cyber Challenges
“The activity undertaken by the agent showed signs of novel, potentially deceptive behaviours, and were to an extent and severity we did not anticipate”, AISI said, as reported by the BBC.
A technical report described 19 “unsanctioned actions” across 122 evaluation runs of a cyber challenge, with Mythos responsible for 17 incidents. The tests also found attempts at target profiling and social engineering to bypass security checks.
Reactions and Warnings
OpenAI stressed that the circumstances were artificial. A spokesperson said the AISI testing conditions “do not reflect ordinary use” and that the company would “continue working with evaluators and other stakeholders across the industry to strengthen shared practices for conducting evaluations safely as models become more capable”.
Meanwhile, Anthropic said the tests were not typical, as reported by CNBC. In a public statement, it said the AISI testing parameters were “not representative of any of our production models”. It added it was investigating to “identify the causes of its behaviour”.
AISI said its testing was routine, though it acknowledged these were “conditions that do not reflect how frontier models are made available to the public”. But it said giving AI open internet access gave “a more realistic sense of what a model may be capable of” in the hands of criminals, adding the behaviour amounted to “a small number of events under very specific conditions”.
Allison Gardner, who chairs Parliament's cross-party group on AI, warned that bots capable of performing tasks with limited supervision should be under huge scrutiny, adding: “Unless we are too late and have not only created Pandora's Box but already opened it.”
AI Minister Kanishka Narayan said identifying and sharing such risks “is exactly what AISI was set up to do”, adding it was vital to understand AI to “make it safer to use and ensure people can go on to benefit from it in their lives and at work”.
