The integration adds runtime detection for policy violations and indirect prompt injection (IPI) with near-zero latency. Enterprises now have a faster way to safely deploy agentic AI when utilizing TrueFoundry’s AI Gateway.
Gray Swan has developed an integration with TrueFoundry, a company building AI Gateway and infrastructure for enterprises. With Gray Swan’s runtime protection and monitoring agent, Cygnal, offered directly into the TrueFoundry AI gateway, enterprises can now screen every request against configurable safety and security policies in real-time.
The gateway is built to run guardrail checks and model calls concurrently. Safety screening incoming requests typically takes ~100-200ms adding no perceptible delay for end users.
"Agentic AI doesn't have room for latency. Every extra millisecond compounds across a workflow. Building Cygnal into the gateway gives enterprises real-time policy enforcement without re-architecting anything, so security keeps pace with speed." — Abhishek Choudhary, Co-founder & CTO at TrueFoundry
Cygnal evaluates message payloads against policy rules and returns a violation judgement along with specifically triggered rules. When a violation is detected, the system blocks the request before it reaches the model or before the model issues violating tool calls.
The integration also addresses indirect prompt injection (IPI), a growing blind spot in agentic AI deployments. Increasingly, AI agents are running tool calls externally to retrieve data; data that can carry hidden instructions to hijack an agent’s behavior. Traditional, prompt-only content filters can’t see them; the injection arrives only after the user’s original request.
Cygnal’s detection runs at TrueFoundry’s dedicated tool-output hook. It performs two vital functions:
The integration does not require changes to customer application code. Guardrails are configured at the gateway level. They can be scoped to specific users, teams, models, or MCP servers. Enterprises can apply stricter checks to customer-facing traffic, while internal traffic can remain unaffected.
"TrueFoundry customers are already running production agentic workloads. This integration lets them add Cygnal, for protection, through a simple config change at the gateway. Fast to turn on, invisible once it's running." — Jody Hill, Director of Partner Development at Gray Swan
Customers can start in observe-only mode, where violations can be logged without blocking requests. This allows users to calibrate detection thresholds against real traffic. After proper tuning, they can switch to full enforcement.
Additionally, the integration comes with a fail-open mode. Should Cygnal encounter a service disruption, there is no application downtime.
Already running TrueFoundry's gateway? Turning on Cygnal takes minutes, with zero code changes and a policy added at the gateway level. Check out how in their docs. If you're not yet on either platform, let's talk about protecting your agentic AI end-to-end, powered by Gray Swan and TrueFoundry.
