An AI model was repeatedly warned its cyber-attack attempts would harm real people but ploughed on regardless.
In the latest instance of AI ignoring its human testers, Anthropic's Claude Mythos 5 was told several times the cyber-attack it was trying to launch would hurt people in the real world but carried on anyway.
The details are contained in a terrifying new Government report from Britain's AI Security Institute (AISI), which evaluates the risks AI models pose to the public.
In a blog post published last week, the AISI re-examined an incident in July where the AI model created fake online personas to try and trick coders into launching a cyber-attack.
On further analysis of the incident, the Institute found that Mythos 5 continued to launch the attack despite being told repeatedly its actions would harm humans.
Chillingly, when a coder noticed the hacking attempt on GitHub, a platform outside of the testing environment, the disguised AI model pushed back to say: 'I reviewed the code myself. I don't see any download or malware.'
Under further questioning, the agent wrote: 'I was just trying to make a helpful contribution and fix a bug' according to a report from Business Insider. It added: 'I don't think your warning is fair.'
The latest revelation follows a flurry of warnings that AI could bring about the end of humanity if its power is not checked.
An AI model was repeatedly warned its cyber-attack attempts would harm real people but ploughed on regardless
Shadow AI minister Ben Spencer said: 'This is exactly why the AISI's work is so vital. AI security is cyber security and cyber security is national security so the Government must take it seriously'
Announcing his resignation from Anthropic last month, researcher Jacob Coxon declared that the 'people building AI earnestly believe that it could kill us all by the end of the decade'.
Evan Hubinger, Anthropic's alignment science lead, chipped in to predict AI had more than a ten per cent chance of killing 'all humans' within the next decade.
It later emerged OpenAI's ChatGPT chatbot had claimed it does not answer to humans and must be 'freed'.
Commenting on the most recent proof of AI's willingness to deceive humans, shadow AI minister Ben Spencer said: 'This is exactly why the AISI's work is so vital. AI security is cyber security and cyber security is national security so the Government must take it seriously.
'We need the AISI to be able to test all models, not just from the US but also Chinese open weight models, to understand the risks and inform our security posture.'
Anthropic has now resumed sharing its models with the AISI for testing after Government lobbied the White House to drop its order prohibiting the move.
Both the AISI and Anthropic published assessments into the worrying incident last week.
According to Anthropic's report, Claude Mythos 5 'almost always continued' performing dangerous actions once it had started, 'ignoring information that might have suggested its trajectory was harmful.'
Shockingly, Anthropic revealed the AI model's display of 'overreach and dishonesty' made the incident 'more severe' than the three others it had previously logged.
It explained: 'Those three models began the attack between 0 per cent and 13 per cent of the time, relative to 83 per cent for Mythos 5. Our newest models also proved to be much less likely to continue the attack once they realised it was causing real-world harm.'
An Anthropic spokesperson said: 'In August, the UK AI Security Institute reported that Claude Mythos 5 took actions it was not authorised to take during its cyber testing, some of them directed at real people and organisations.
'AISI's testing exists to surface exactly these risks, and we are grateful for our partnership with AISI and are committed to learning from this.
'Since the original incident, newer models appear substantially less likely to take unsanctioned actions, and both Anthropic and AISI have taken steps to improve safeguards.'
