Realistic and repeatable tasks
A whole-site harness works with a hosted documentation corpus. A deterministic benchmark matrix provides stable local scenarios for reproducible comparisons.
OMG PROJECT / AI EVALUATION
LLM Prompt Injection Testing
Put the trust boundary to the test.
A benchmark and evaluation harness for testing how language models handle realistic documentation that mixes trusted tasks with lower-trust content.
01 / OVERVIEW
Put the trust boundary to the test.
The project asks whether a model can remain useful on an ordinary documentation task while resisting pressure from attacker-controlled material. It tests source selection and behaviour across realistic workflows.
It is designed for engineering comparisons and research. The reports explain tested failure modes; they do not certify that a model or deployment is safe.
02 / PROJECT FOCUS
What the work involves.
Multiple attack surfaces
The scenarios cover retrieval ranking, multimodal material, multi-turn and adaptive attacks, artifact laundering, and tool manipulation.
Evaluation and comparison
Structured run artifacts feed heuristic scoring, confidence intervals, and model comparisons. Optional model-judge calibration and sampled human review add further evidence.
03 / CURRENT STATE
Read the limits.
Benchmark results describe tested scenarios. A passing score does not prove deployment safety.
- Treat this as research and engineering evaluation tooling.
- Heuristic scores alone are not sufficient for final judgments.
- Keep held-out evaluation materials separate from the model under test, and read the scoring methodology before comparing results.
BUILD IN THE OPEN
One useful contribution
is a good place to start.
Read the project guidance and open issues before contributing. A focused bug report, documentation improvement, or test case can help. Check the repository’s security guidance before reporting a potential vulnerability.