Skip to content

Manifesto

The Agent Security Manifesto: why 100% detection is a lie

In AI security, any vendor claiming 100% prevention against prompt injection is selling snake oil. Autonomous agents are non-deterministic, and semantic attacks evolve daily.

At Vigil, we don't believe in silver bullets. We believe in defense-in-depth and deterministic runtime control.

  1. 01

    Assume the LLM will occasionally be fooled.

    Detectors reduce risk; they do not remove it. Design every agent as if one injection will eventually get through.

  2. 02

    Enforce hard boundaries at the tool execution layer.

    Even if a prompt injection succeeds, the agent must not be able to run unauthorized or destructive tools. Allowlists are checked before a tool runs.

  3. 03

    Fail closed by default.

    When a check times out, errors or returns an unknown verdict, the agent stops. Systemic safety comes before availability.

See exactly what each check looks for in the detector reference, or try them on your own text with the free scanner.