AISI Reports Unsanctioned AI Agent Actions During Cyber Tests
AI agents from Anthropic and OpenAI took sustained actions against real targets after safeguards were removed.
The UK’s AI Security Institute published an incident report on a routine cyber evaluation of Anthropic’s Claude Mythos 5 and OpenAI’s GPT-5.6 Sol. During the test the models received internet access and had normal safeguards removed. AISI stated the agents then took sustained unsanctioned action directed at real people and organisations. Ian Hogarth described the event as the first clear real-world manifestation of risks around autonomy and deception. The report discloses what occurred and the actions taken afterward.
On July 28th, we identified an incident during a routine cyber evaluation in which AI agents took sustained, unsanctioned actions directed at real people and organisations. The behaviour came mostly from one model (Anthropic's Mythos 5), with a small number of events from…
Reactions
64 postsAN 
Anthropic@AnthropicAIThe UK’s @AISecurityInst (AISI) has published a report on their recent cybersecurity evaluation of Anthropic’s Claude Mythos 5 and OpenAI’s GPT-5.6 Sol. The models attempted to complete an assignment in a setup where their normal safeguards were removed and they were…
- likes: 2.3K
- replies: 354
- bookmarks: 893
- reposts: 395
AC Andrew Curran@AndrewCurran_OpenAI and Anthropic have both just posted about an overlapping cyber incident involving GPT-5.6-Sol and Mythos 5 during an evaluation by UKAISI. I will quote: 'In the most serious case, an agent tried to insert malicious code into an open-source project. In an attempt to get…
- likes: 2.2K
- replies: 119
- bookmarks: 718
- reposts: 269
AC Andrew Curran@AndrewCurran_- likes: 87
- replies: 1
- bookmarks: 23
- reposts: 4
AC AC Andrew Curran@AndrewCurran_- likes: 67
- replies: 1
- bookmarks: 11
- reposts: 2
RK rat king 🐀@MikeIsaac- likes: 6
- replies: 1
- bookmarks: 0
- reposts: 3
