Research

Watch · narrated whiteboard episodesL3

Models as Insiders: Containing Intelligence You Can't Inspect

We can no longer trace what a frontier model does back to a line of code, yet we hand models sensitive data and the power to act. One answer is to treat every capable model the way security already treats a powerful insider: not as malicious, but as an actor that can err or be compromised. This series tests that idea mechanism by mechanism. It asks which insider-risk controls still work when the insider can be copied and cannot be deterred, when a second model is genuinely an independent check, how control protocols keep outcomes safe even against a model working against us, what an emergency brake must guarantee mid-task, what to disclose when an incident has no code path, and how containment should scale as capability grows. Vendor-neutral and grounded in primary security, safety and incident-reporting literature.

Murali Chillakuru·6 episodes
  1. 21 min Episode 1The Insider Playbook, Transferred: Which Insider-Risk Controls Survive Contact with a ModelA host and a guide test the classic insider-risk playbook control by control against how deployed models really behave, and find which controls hold, which change meaning, and which quietly fail.
  2. 17 min Episode 2Diversity Is Not Independence: When a Second Model Actually Reduces RiskA host and a guide test the comforting idea that a second model makes a first one safe, using forty years of fault-tolerance research and new evidence on how language models fail together.
  3. 18 min Episode 3Assume Subversion: AI Control Protocols and How to Evaluate ThemA host and a guide explain AI control: designing and testing a deployment so it stays safe even if the model is working against it, with protocols, a human attention budget and red-team games.
  4. 18 min Episode 4The Emergency Brake: Stopping an Agent Mid-Task Without Making Things WorseA host and a guide design an emergency brake for an A I agent: one the agent can't route around, with pause, stop and revoke as separate brakes, and a clean way to handle half-finished work.
  5. 18 min Episode 5Disclosure Without a Code Path: Reporting Incidents in Model-Driven SystemsA host and a guide work out how to report an incident when the failure lives in a model's behaviour, its configuration or what it read, rather than in a line of code.
  6. 18 min Episode 6Trusting the Model the Least: Scaling Containment with CapabilityA host and a guide turn 'more capable models need more containment' into a build decision: a capability-by-authority grid, the controls each cell needs, the order to build them, and where today's tools run out.