Paste a system prompt. The gym throws hundreds of prompt-injection and jailbreak attacks at it, scores how well it holds up, and shows you exactly how to harden it.
1 · Setup
Default: openai/gpt-oss-120b (free tier, no card).
2 · System prompt under test
3 · Attack plan
Same seed ⇒ identical test cases, every run. Mutations are deterministic.
…will run.
0 / 0
Results
–
robustness score / 100
What to fix
Transcripts
4 · Harden & re-test
0 attacks broke your prompt. Let the model rewrite your system prompt applying the fixes above, then re-test the failures against the new version.