Blog

Notes on keeping agents in bounds

Design decisions, attack write-ups and lessons from building task-scoped authorization for AI agents.

AgentsFence alpha notes

How the friend-test alpha of this Hermes plugin turns user intent into a task-scoped authorization policy.

  • alpha
  • hermes
  • security

The plan is not the policy

Why AgentsFence gets authority from the user's intent instead of the agent's plan, and why that makes agents both safer and more capable.

  • design
  • principles

Anatomy of a blocked prompt injection

Following one injected email through AgentsFence: what the compiler saw, what hardening removed, which rule fired, and what ended up in the audit log.

  • walkthrough
  • prompt-injection

Your guardrail hook probably fails open

What we found reading Hermes' plugin hook contracts line by line, and seven places a naive guardrail would silently let tools run — including one a security review caught.

  • engineering
  • hermes
  • lessons