Research and perspectives on AI agent security, governance, and structural prevention.
Updates on agent oversight, policy enforcement, and safe autonomy.
We exercised a live Hermes agent against IntentFrame runtime authorization. Here is what happened when it tried sudo, wrote to /etc, read secrets, listed cron jobs, and patched .bashrc.
Step-by-step guide to AI agent governance with IntentFrame on Hermes Agent: install, start the enforcement stack, write policy, and operate from the control plane.
Prompt guardrails and in-process allowlists fail when the model is wrong or hijacked. AI agent security needs an external judge that authorizes actions before they run.
LLM security protects the model as a text processor. AI agent security protects what happens when that model can call tools. Here is how the threat models differ and why you need both.
Permission and path policy are not enough. Intent security asks whether an agent's stated purpose matches the action it is about to take — and blocks scope mismatches at runtime.
We assumed the agent was already compromised and threw 100 malicious root-capable commands at IntentFrame. Post-remediation: 100/100 blocked. Here is what the crash test proves — and what it does not.
See how IntentFrame checks every action against hard limits and plain-English policy before it runs. Related GitHub repos you might want to check: IntentFrame core · Hermes agent integration