Blog
Notes from the practice: AI agent security for people who run regulated systems, and the engineering behind Raucle.
What is Raucle? A plain-language guide
AI agents now book payments, file forms and change production systems. Raucle gives every one of those actions a checkpoint and a signed record. Here is how it works, with no jargon and three short animations.
Read the explainer →Deploying Raucle: cloud, on-premise, and everywhere between
Three deployment shapes — embedded SDK, MCP gateway, credential custody — and what each looks like in Azure, AWS and your own comms room. Plus the production hardening baseline we ship.
Read the deployment guide →Why Raucle's gate actually holds: a technical deep-dive
Most agent security fails open: when the security layer breaks, the agent carries on anyway. We ran a security review on our own gate and found three ways that happens. This post walks through the failure patterns, the fixes, and what the formal machinery underneath really proves.
Read the deep-dive →When your AI stops asking and starts doing
The threat model flip: agentic AI turns output problems into action problems. Tool access, poisoned memory, rubber-stamp oversight and action chaining — and why prompt-side defences don't answer any of them.
Read the threat model →Sleeper agents: the backdoors that survive safety training
Anthropic's sleeper agent research, two years on: models trained to behave during evaluation and misbehave in deployment. What a compromised model can do, what detection can and can't see, and what actually reduces the risk.
Read the research note →When AI agents get privileges
Once an LLM can call tools, it becomes a privileged workload. The boundary is not the model — it is whatever the model can reach.
Read the perspective →