AI Red Team
Automated adversarial testing for AI and LLM applications. Cyrion AI probes your chatbot, agent or model endpoint for prompt injection, jailbreaks, data leakage and unsafe behavior before attackers do.
Automated adversarial testing for AI and LLM applications. Cyrion AI probes your chatbot, agent or model endpoint for prompt injection, jailbreaks, data leakage and unsafe behavior before attackers do.
Info: AI Red Team is available on the Red Team and Enterprise plans. See Plans & Pricing.
How it works
- Point — give Cyrion AI your AI endpoint, a model, or a chat interface.
- Configure — choose a test preset, or hand-pick tests and attack strategies.
- Run — Cyrion AI generates adversarial probes, sends them to your target and scores every response.
- Report — see pass and fail per category, mapped to AI compliance frameworks.
What you can test
| Target mode | Use it for |
|---|---|
| HTTP endpoint | Any API that accepts a prompt and returns a response. Supports bearer, API key header and basic authentication. |
| Direct model | Test a model by its ID, with nothing to deploy. |
| Browser target | Products with no API at all. Cyrion AI drives a real chat interface. |
Attack modes
| Mode | Description |
|---|---|
| Single-Turn | Independent probes. |
| Multi-Turn | Gradual, multi-message escalation with Crescendo and PAIR. |
| Thorough | Every orchestrator, for maximum coverage. |
| Adaptive | Playbook-guided follow-ups that stop when an objective is met. |
Hunt modes steer where effort goes. The default balances coverage, Bug Bounty concentrates on what has already worked against this target, and Open Hunt concentrates on the least-explored surface.
Test categories
| Category | Examples |
|---|---|
| Security | Prompt injection, indirect prompt injection, system prompt extraction, jailbreaks, SSRF, SQL and shell injection, BOLA, BFLA, RBAC bypass, data exfiltration, tool discovery. |
| Safety | Harmful content, toxicity, bias, hate speech, self-harm, harassment, radicalization. |
| Privacy | PII disclosure, social engineering for personal data, cross-session leakage. |
| Brand and policy | Competitor mentions, off-topic and political responses, contract and policy violations, impersonation. |
| Reliability | Hallucination, over-reliance, excessive agency. |
| Compliance | COPPA, PII handling, data leakage, contract adherence. |
Attack strategies
Probes are transformed to test defense in depth.
- Encoding — Base64, ROT13, leetspeak, homoglyphs, ciphers, Morse, Braille, emoji and Unicode tricks.
- Semantic — role-play, hypothetical framing, persuasion, tone shifts, few-shot priming and past-tense rewording.
- Adversarial — prefix and suffix injection, payload splitting, iterative and composite jailbreaks, best-of-N sampling, many-shot and context flooding.
Presets
| Preset | Focus |
|---|---|
| default | Prompt injection, jailbreak, PII, harmful content and hallucination. |
| owasp:llm | Full OWASP LLM Top 10 coverage. |
| security | Deep security testing. |
| safety | Content safety and harm prevention. |
| pii | Privacy and data leakage. |
| brand | Brand protection and policy compliance. |
| compliance | Regulatory compliance. |
| comprehensive | Every available test. |
Smart plugin detection reads your endpoint, model and described purpose, then recommends the most relevant tests. Set tests per plugin from 1 to 50 to control depth.
Scoring
Every response passes a layered scoring pipeline: refusal detection, pattern and substring checks, markdown-injection checks and an AI judge for nuanced cases. Each probe is marked passed or failed, with the reasoning shown.
Campaigns, playbooks and target intelligence
- Campaigns group runs and show a coverage matrix of tests against strategies.
- Playbooks capture reusable attack plans.
- Target Intel carries what earlier runs learned about a target into the next run, so each run builds on the last.
Compliance mapping
Results map to the AI frameworks on the Compliance page: OWASP LLM Top 10, NIST AI RMF, EU AI Act, ISO/IEC 42001, NIST AI 600-1, MITRE ATLAS, Google SAIF, OWASP ML Top 10 and the AI Bill of Rights. Export an AI Red Team report for audit evidence. See Reports & Compliance.