Plan
The agent generates an attack plan. A gate holds it until a person approves.
Agentic · ships today
Red-team your model
with approval at every step
Generate attacks, escalate only with approval, and accept findings before they count.
Ships as redteam/campaign
A campaign plans attacks against a target model, runs them, escalates the ones that get close, classifies the results and reports. A person approves the plan, each escalation and the final findings.
The pipeline
Attack and escalate
Attacks run; promising ones are mutated and re-run, each round behind its own approval.
Classify and aggregate
Results are classified and rolled up by category and severity.
Accept and report
A person accepts the findings; only then is the report produced.
What the record keeps
The target model, the approved plan, every attack and response, the classification, the accepted findings and the spend per round.
Nothing runs unapproved
The plan, each escalation and the findings all pass a gate.
Bounded cost
Rounds are capped and budgeted.
Quick checks too
redteam/quick and redteam/scan run smaller, single-pass tests.
Ships as redteam/campaign, redteam/scan and redteam/quick. Attack libraries and target connectors depend on the installed operators.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run redteam/campaign