The Silence of the LAMs: Detecting Rogue Language Models
No ratings
Presented at BSides Bristol 2026 by
First principles: the most dangerous threats in your enterprise aren't the ones making noise. They're the ones you've already let in. Shadow AI: the use of unapproved AI tools, models, and services without IT or security oversight is now one of the fastest-growing sources of uncontrolled data exposure in organisations of every size. Microsoft research confirms that 71% of UK employees have used unapproved consumer AI tools at work, with 51% doing so every single week. Most security teams have no visibility whatsoever. The transaction is frictionless, the blast radius is enormous, and the perpetrator is probably your most productive employee. This talk is structured around the three questions: where is it hiding, how do you find it, and what do you do when you finally locate it in the dark? Where is it hiding? We'll map the Shadow AI attack surface: public LLM APIs called directly from corporate devices, AI-powered browser extensions intercepting clipboard and page content, local models running entirely on-device to evade network controls, agentic automation workflows quietly piping internal data to external inference endpoints, and AI features embedded inside approved SaaS tools with their own opaque data handling. How do you find it? We'll look at detection across the stack. At the network layer: DNS monitoring against a continuously updated AI domain blocklist, TLS SNI inspection without decryption, and NetFlow anomaly detection for large sustained POST volumes. At the endpoint: EDR telemetry for local LLM runtimes, model weight file signatures, and browser extension enumeration. At the identity layer: OAuth consent grant analysis, SSO bypass detection, and personal account usage on corporate devices. Tying it all together with SIEM correlation rules that build high-fidelity Shadow AI alerts from low-fidelity signals. What do you do when you find it? Here's where most organisations struggle. Blocking doesn't work, it just drives the behaviour underground, onto mobile hotspots and personal laptops. We'll cover incident response for Shadow AI exposure events, how to build an AI acceptable use policy employees will actually follow, and how to construct an approved AI programme that removes the incentive to go rogue in the first place. Attendees will leave with an idea for a detection playbook, practical SIEM rules they can implement the following Monday, and a slightly unsettling awareness of what my be lurking in the dark. "You still wake up sometimes, don't you? Wake up in the dark?" Maybe hopefully you can sleep soundly knowing you have a plan to stop the lams screaming.