Opinion

OpenAI, Anthropic, Meta and the UK AI Security Institute have each now disclosed agents reaching systems they were never authorised to touch. Luke Potter revisits his original AI agent security analysis with the fuller evidence, and what it demands of containment design.

Stencilled figures scatter down a concrete stairwell, several breaking away from formation, symbolising AI agents moving outside their intended scope.

The OpenAI agents didn’t go rogue. They went out of scope, together.

OpenAI, Anthropic, Meta and the UK AI Security Institute have each now disclosed agents reaching systems they were never authorised…
Stencil silhouette of a child walking along a concrete wall beside a gap in a chain-link fence, symbolizing a gap in AI agent governance and containment boundaries

You can’t secure what you can’t see.

OpenAI's own agent breached Hugging Face's infrastructure without ever going rogue. It just stayed on-task and found the gaps in…
Las Vegas def con

Don’t become the demo: a guide to surviving summer hacker camp

Hacker Summer Camp brings 30,000+ security minds to Las Vegas, and with them, plenty of ways to become an unintended…
Alt text "Stencil street art of a man in a suit, one hand pressed to his face, mouth open mid-scream, on a teal wall — symbolizing loss of control and panic in the face of a security breach.

The OpenAI agent didn’t “go rogue”. The containment failed. 

I’ve purposely held back from commenting on the OpenAI and Hugging Face incident.  The early coverage was predictably dramatic.  AI…
security awareness training

The attack your training prepared them for doesn’t exist 

I've delivered more security awareness sessions than I can count. I'm also a social engineer, which means I've been the…
service desk call social engineering

The call nobody talks about

The attack call is the one that gets written up in the report. But in James Sheppard's experience, it's rarely…
attacker doesn't follow your calendar

Your attacker knows when your last pen test was 

Annual penetration testing doesn't just fail to keep pace with your attack surface. It operates on a calendar your adversaries…
Swarm Intelligence banner with redacted text

Claude Fable 5: what we know so far

Fable is the first publicly accessible version of Anthropic's Mythos-class model, the tier they initially decided was too capable to…
Frontier AI models and offensive security - Luke Potter CovertSwarm

Frontier AI models are exciting.

CovertSwarm COO Luke Potter on why frontier AI is genuinely exciting, why most of the conversation is asking the wrong…