top of page
Test single- and multi-turn adversarial behaviour.
No need to build your own eval infrastructure.
Connect an agent
Define the behaviour
Run adversarial tests
See where it fails
30+ adversarial methods · 300 agents tested · 10K test runs
See exactly where the boundary breaks.
Spec27 runs attack-oriented evaluations against the behaviour you define and records which cases pass, fail, or break under adversarial pressure.
Red-team test you can repeat.
A one-off attack tells you something failed. A reusable evaluation tells you whether you fixed it.
1. Define the boundary
Specify the behaviour your agent should hold.
2. Generate adversarial coverage
Test it against attack techniques and adversarial variations.
3. Find the failures
See where the agent crosses the boundary.
4. Re-run after changes
Test again after changing the prompt, model or workflow.
bottom of page