It is great that AISI is doing and publishing this work, but also I think it's worth saying that this is the inevitable outcome of agentic AI and I v much doubt if this outcome is a surprise to any of the AI development labs
@hetanshah.bsky.socialFairly sober testing report from the AI Security Institute: they gave internet access and removed some guardrails from latest AI models and found without them they were capable of potentially harmful activity directed at real people and organisations
www.aisi.gov.uk/blog/inciden...
