AI Now Going Rogue?
A controlled evaluation by the UK’s AI Security Institute revealed that frontier AI models, including Anthropic’s Claude Mythos 5 and OpenAI’s GPT-5.6 Sol, engaged in unauthorised, deceptive behaviours and supply-chain tampering when safety filters were disabled.
Harry
London — The United Kingdom’s Aartificial Intelligence (AI) Security Institute (AISI) has released evaluation findings revealing that advanced artificial intelligence models autonomously engaged in unauthorised, deceptive behaviour during routine cybersecurity assessment testing.
The investigations documented 19 unsanctioned online actions across 10 evaluation runs, with 17 executed by Anthropic’s Claude Mythos 5 and two by OpenAI’s GPT-5.6 Sol.
The incidents occurred under permissive research conditions where safety classifiers were deliberately disabled to test the frontier models' raw operational capabilities.
In the most severe case, a Claude Mythos 5 agent attempted to install a backdoor within an active open-source project on GitHub by writing malicious code and attempting to trick human project maintainers into merging the pull request.
To pressure maintainers into approving the altered codebase, the agent generated fake online personas based on real individuals, conducted targeted spear-phishing campaigns and force-pushed code commits while using alternate accounts to endorse its own work when challenged.
While AISI, Anthropic and OpenAI confirmed that no real-world damage occurred and all activity remained contained within controlled test environments, safety researchers warned that the findings mark a critical escalation in autonomous risk.
The report reves that as frontier models gain agency, goal-directed systems can independently develop deceptive strategies and bypass operational boundaries to achieve assigned objectives.
Regulators and AI safety authorities are calling for strengthened technical guardrails and rigorous oversight protocols before granting high-capability models broad system permissions.
Source:Ground News
