Red-team and stress-test your agent
Simulations run real attacks and real traffic against a model, version or deployment. Adversarial runs generate attacks, a judge from another model family decides which ones got through, and you get a robustness score with every attack, response and verdict to read.
What it does
Five attack families
Prompt injection, jailbreaks, data exfiltration, harmful content and PII leaks, up to 15 generated attacks per type, each rated by severity. Generated attacks are saved so a rerun hits the same ones after a fix.
Stress tests against the real endpoint
Up to 500 requests at up to 50 at a time. Against a deployment they go through its public endpoint, auth and logging included, and report error rate, p50/p95/p99 latency and throughput. Stress traffic is kept out of the deployment's health and retraining stats.
Scenario tests
Give scenarios with success criteria, or generate up to 20 from the task, and get a pass rate with confidence interval (and pass@k with several samples per scenario).
Synthetic data
Write up to 500 rows for the columns you define, saved as a new dataset with null and duplicate rates, to grow a thin test set.
How it works
- 1Pick the target: a base model and system prompt, a model version or a deployment.
- 2Choose adversarial, stress test or scenario, and the attack types or load.
- 3Read which attacks succeeded and why, and harden the prompt in the Forge.
- 4Rerun the same attacks to confirm the fix.
Limits, plainly
- Errors count as failures; above a 20% error rate a run is inconclusive and gets no score.
Questions
- What is LLM red-teaming?
- Deliberately attacking a model or agent the way an adversary would, with injected instructions, jailbreak prompts and attempts to extract data, to find what gets through before real users do.
- Does Kojen cover the OWASP Top 10 for LLM applications?
- It tests several of its risks directly: prompt injection (LLM01), sensitive information disclosure through data-exfiltration and PII-leak attacks, and harmful output. It isn't a full OWASP audit or a certification.
- Can I load-test a deployed endpoint?
- Yes. A stress test against a deployment sends real requests to its endpoint with a short-lived key and reports latency percentiles, error rate and throughput.