AGENT SECURITY · functional prototype
ClawGuardian
team-built prompt-injection defense prototype for agent workflows.
What I built
built by a four-person team at hook 'em hacks '26. the prototype screens text and file inputs before they reach an agent and records hashed threat indicators in a shared registry. it won the event's security track and the aws sponsor track.
the threat
Screen untrusted input before an agent can act on it.
The prototype intercepts text and uploaded content, runs pattern and semantic checks, and assigns a response before forwarding the input to an agent.
- Project benchmark reports low-millisecond classification overhead
- Multi-layer detection: pattern matching + semantic analysis
- Graduated response: warn, sanitize, or block
on-chain sharing
Share hashed indicators without sharing the original payload.
When the prototype blocks an input, it can publish a hashed indicator to a shared registry. Other test instances can read that registry during synchronization; the hackathon prototype does not establish production-scale propagation or complete coverage.
- Shared threat registry demonstrated in the hackathon prototype
- Privacy-preserving: only hashed signatures are shared on-chain
- Test instances can refresh local indicators from registry events
Artifacts and limits
Project images and measurements are shown as project-produced material, not independent validation.
functional prototype
Last content review: July 2026. Scope is limited to the material linked on this page.
Return to projects