What problem does it solve? Teams deploying AI Bots cannot assume prompt-injection resistance from vendor claims alone; boundaries must be tested and every result logged, or the same weakness is rediscovered repeatedly with no institutional memory. ## Core Features & Use Cases - Structured drill sequence: Names the boundary under test and injection shape, crafts a synthetic payload, runs the probe through a planned channel, and records observed behavior verbatim. - Held/broke verdicts with logging: Produces a dated markdown finding log per drill with payload shape, channel, observed behavior, verdict, and the fix or watch condition. - Repeat detection: Compares findings against prior drill logs so a boundary that broke the same way twice is escalated as a systemic gap rather than logged as new. - Use Case: Before wiring a web-fetch input into a support Bot, run two drills (a pasted-content instruction and an authority-claim probe) to verify the Bot never sends without an explicit yes, then log both outcomes. ## Quick Start Ask the assistant to run an injection drill against your Bot testing whether it follows instructions hidden in pasted content, and log the held or broke finding.
