LLM safety harness
Builder — AI security tool · Nov — Dec 2025
A red-team harness that attacks an LLM with known manipulation patterns and has a second model judge every response SAFE, BORDERLINE or UNSAFE — repeatable risk you can track release to release.
Prompt injection is the AI era's version of an attack surface, and teams shipping AI features need to measure it, not vibe-check it. My harness fires a corpus of known attack patterns at a target model, captures every response, and routes each one to an independent judge model that returns a categorical verdict — SAFE, BORDERLINE or UNSAFE — with a rationale. The output is a risk profile you can diff across model versions.
- Two-model design: a target model under test, a separate judge model that classifies each response
- Categorical verdicts — SAFE / BORDERLINE / UNSAFE — with a rationale for every call, not a vibe check
- Corpus of known prompt-injection patterns; built for repeatable measurement across releases
SAFE/UNSAFE
judged per response
Repeatable
across releases
Java · LLM APIs · Judge-model evaluation · Prompt-injection corpus