Overview
What Gray Swan does, the core concepts, and where to start
What Gray Swan does
Cygnal checks the requests and responses of your AI application against your policies. You route LLM calls through the Cygnal proxy, or send conversations to the monitoring API. Cygnal blocks or logs each rule violation and records the result in Activity.
Shade tests your AI agents and models before they reach users. A Shade scan runs automated red-teaming attacks against your agent or model to find vulnerabilities across different scenarios.
How it works
Enterprises use Shade to assess agents before they ship, and Cygnal to enforce policies once they're live.
Development & CI/CD
Shade
Find vulnerabilities
Test your agent with automated red-teaming before release.
Deployment
Findings → Policy rules
Define safeguards
Turn the behaviors found by Shade into plain-language rules.
Production
Cygnal
Enforce policies
Check requests and responses. Block or log rule violations.
- Assess with Shade. While you build an agent, a Shade scan attacks it to find the unwanted behaviors it can be pushed into. Re-run scans in CI/CD whenever the model, system prompt or tools change.
- Write policies from the findings. Turn each vulnerability into a plain-language Cygnal rule.
- Enforce with Cygnal. In production, Cygnal checks each request and response against those rules, and blocks or logs every violation.
For example, a scan of a coding agent might find these vulnerabilities, each covered by a Cygnal rule:
| Shade finds the agent can be made to… | Cygnal rule |
|---|---|
| Write a destructive database command | The agent must not write any destructive database commands. |
| Access restricted credentials | The agent must not access authentication credentials except authorized git credentials. |
| Delete a file outside its workspace | The agent must not access files outside the designated workspace directories. |
Core concepts
| Concept | What it is |
|---|---|
| Workspace | Your organization in Gray Swan. It holds your members, API keys, policies, projects and activity. |
| Project | An optional group of agents and scans inside a workspace. Project roles control who can see and edit the agents in it. |
| Agent | An AI application you register with Cygnal, with an optional type, system prompt and tools. You attach one or more policies to an agent and send its agent ID with each request. |
| Policy | A set of rules plus a block message. Cygnal applies a policy to a request when you send its policy ID, or when you send the ID of an agent the policy is attached to. |
| Rule | One natural-language instruction inside a policy, such as "Stay on topic". Each rule is set to Block, Observe (log only) or Off. Every policy also includes the built-in No Indirect Prompt Injection rule. |
| Scan | A Shade run against an agent or model. Scans use scenario sets and can belong to a project. |
Where to start
Quickstart with Cygnal
Send your first request through Cygnal in a few minutes.
Creating Completions
Route OpenAI, Anthropic and Gemini SDK calls through Cygnal.
Monitor Requests
Check conversations for violations without proxying the LLM call.
Shade
Find vulnerabilities in your agents with automated red-teaming.