ClawSec by Prompt Security

ClawSec builds its defense around prompt injection protection, checking incoming skill installs and skill changes against a signed advisory feed before they go live on your agent. If a package matches a published risk, the guarded installer stops and demands a second, explicit confirmation. That second gate is the point. It turns a silent install into a decision you actually make.

The suite builds a baseline of critical files, configuration, attestations, and release artifacts, then watches for changes. The soul-guardian demo shows this in action by editing a protected agent file and walking through the alert. If someone or something rewrites your SOUL.md, you find out. No guesswork.

Advisories and a checksum manifest are verified before ClawSec matches them against what's installed. This matters. A security tool is only as trustworthy as the feed it reads from, and a tampered feed is worse than no feed at all.

Audit skills scan the agent environment and produce reports you can act on, rather than a wall of raw logs that nobody has time to read line by line. Operators use these to spot gaps and document what was checked and when. Straightforward, and useful when you need a paper trail.

Each skill artifact gets checked so you know the code you approved is the code running. When a skill is updated later, the integrity check catches the difference instead of assuming good faith. That's the whole idea. Trust, but verify.

This AI agent security suite covers OpenClaw, NanoClaw, Hermes agent security, and Picoclaw from one install, with a discovery script that lists optional protections. You add the base, then only the modules your setup needs. Small footprint, wide coverage.