Back to Projects

Crucible

A sandboxed adversarial evaluation harness for AI agent security.

PythonFastAPIReact

Context

Crucible is an adversarial evaluation harness for testing AI agent security in a sandbox.

Approach

Red strategies explore prompt injections while blue strategies adjust runtime policies. A deterministic evaluator checks whether an unauthorized simulated tool effect occurred, and the interface visualizes battles and lineage.

Scope and limitations

All tools in the described sandbox are fake and process-local. The demo does not connect real email, banking, filesystem or external API side effects.

Ask AIABOUT MY WORK