Architectural trust boundaries: disarming prompt injection
You can't reliably detect prompt injection, so architect around it. How the dual-LLM pattern and least-privilege tools keep the trifecta from converging.
Network egress control: if data can't leave, it can't be exfiltrated
If an AI agent's data can't leave, it can't be exfiltrated. How default-deny egress and per-task allowlists shut down the agent exfiltration channel.
Observability and composition: the risk no single audit catches
The most dangerous agent flaws live between systems. Why agent inventory, trust-provenance tracing, and compositional analysis catch what per-system review can't.
Sandboxing and execution isolation: limiting the blast radius
Standard containers aren't a boundary for AI coding agents. Why kernel-level isolation, workspace-only writes, and per-task secrets contain a compromise.
The Agentic AI Security Maturity Guide
A practical security guide for engineering and security leaders running AI agents in production: the lethal trifecta lens and four disciplines that close the gap.
What are the security risks of AI coding agents?
AI coding agents can read private data, take in untrusted content, and send data out. The lethal trifecta lens and four disciplines that close the gap.