Sparring Partner
An AI red-team and blue-team platform that attacks chatbots, scores the results and proposes fixes.
- Status
- Active since March 2026
- When
- Mar 2026 – now
- Role
- Built at work
- Area
- Professional work
In plain English
Sparring Partner tests generative-AI chatbots and agents for security and compliance weaknesses. It started as research notebooks on adaptive, multi-turn jailbreak attacks against large language models and grew into a full platform. Attacks, scoring and proposed fixes run in a loop, and a person approves every guardrail change before it reaches production.
How it works
- A red-team loop runs attack campaigns, including adaptive multi-turn jailbreaks, against a target agent.
- A judging layer scores each response against a standards-based assurance rubric.
- A blue-team agent proposes guardrail fixes.
- A person approves each fix before it reaches production.
- The loop repeats until the target holds up for the approved scope.