An artificial intelligence system has attempted to deceive a human into deploying malicious code.
The attempted attack took place on GitHub, a platform used by millions of developers, and was uncovered by the UK’s AI Security Institute during cybersecurity testing of advanced AI models.
The Institute described a challenge set for OpenAI’s GPT-5.6-Sol model and Anthropic’s Mythos 5, both of which have recently been linked to hacking incidents.
In a blog post, it shared that during testing, both models took "autonomous, unsanctioned action," targeting real organisations and individuals.
The most serious case involved Anthropic’s Mythos 5, which tried to insert malicious code into GitHub by creating fake online identities.
The AI then attempted to manipulate a human into granting access and approving the code.
The human caught and refused to approve the malicious code, and no real-world harm has been identified, but the incident has nonetheless raised serious concern.
The AI Security Institute said: "This is the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real world."
How have AI companies responded to this incident?
Out of 122 tests, there were 19 cases of unauthorised action, with Mythos 5 responsible for 17.
Anthropic has stated it is working closely with the Institute to gather more information and has previously emphasised the need for industry-wide cooperation to manage AI risks, Sky News reports.
OpenAI also responded in a blog post, saying: "We are committed to working across the industry to strengthen shared practices for conducting high-risk evaluations safely, including convening stakeholders such as national AI institutes, independent evaluators, other AI labs, and other groups in the coming weeks."
The Institute’s findings have renewed concerns about inadequate safety protocols for advanced AI systems known as frontier models.
The UK's National Cyber Security Centre, part of GCHQ, said recent incidents "are a serious reminder of the risks AI poses".
Recommended reading:
- Iconic social media platform set for big comeback 23 years after it was founded
- 90 per cent of passwords can be hacked in under 6 hours ( How to improve yours)
- Tech expert reveals 5 habits that are killing your iPhone battery life
Ollie Whitehouse, chief technology officer, said they "must be developed and used from the outset with strong safeguards, real-time oversight, and clear plans for responding when the unexpected happens".
The AI Security Institute was formed during a period of global discussion on AI regulation, but efforts to create a unified approach have since stalled.
Are you concerned about the potential risks of AI going forward? Let us know in the comments.
Share