AI agents are execution environments. Treat them like one.
Linear studies how agents fail once they have real tools, credentials, and filesystem access. We build Rampart to stop unsafe actions before they run, and Snare to catch credential misuse when prevention is not enough.
Incident analysis and threat models for agentic systems.
Rampart enforces policy before commands, files, and tools execute.
Snare plants canaries where compromised agents look first.
Recent Writing
All posts →Tools
Two layers, same threat model: prevent what you can, detect what escapes.
Firewall for AI agents. Intercepts at the OS layer — blocks exfiltration, restricts tool calls, enforces policy before execution.
Canary tokens for AI agents. Plant tripwires in secrets and files. Get alerted the moment an agent touches something it shouldn't.
About
Linear is a security research lab run by Trevor (GitHub: peg). We focus on the security properties of AI agent systems — how they fail, how they get exploited, and what defense looks like at the OS layer. Rampart and Snare are the tools that came out of that work.