TrustGraph, a new open-source policy engine, has launched to solve a critical security gap in AI agent architectures: the inability of standard permission systems to track where data goes after an agent reads it. While traditional Access Control Lists (ACLs) effectively gatekeep what an agent can access, they fail to enforce restrictions on derived outputs. TrustGraph introduces a deterministic, provenance-based model that ensures if an agent reads sensitive data, any output derived from it is automatically restricted to the same audience, closing the loop on silent data exfiltration.
The Invariant Labs Precedent
The project was motivated by a specific vulnerability highlighted by Invariant Labs in May 2025. In that incident, a coding agent was permitted to read private repositories and subsequently open a pull request on a public repository containing the sensitive findings. Each step was individually authorized, but the composite action leaked data because no system checked that the public output derived from the private input. TrustGraph addresses this by enforcing a strict rule: the readership of an output is the intersection of the readership of all its inputs.
Deterministic Policy Enforcement
Unlike LLM-based guardrails, TrustGraph relies on a deterministic kernel with no model in the decision path. It uses a YAML-based policy file to define groups, agents, sources, and actions. When an agent attempts an action, the system calculates the allowed audience based on the provenance of the data being sent. If the target audience includes users not authorized to read the source data, the action is denied or flagged for approval. This approach eliminates the probabilistic nature of AI-driven security checks, offering verifiable guarantees.
Integration and Runtime Architecture
The tool provides a Rust-based CLI and a gateway server that sits between agents and their tools (such as MCP servers). The gateway holds the credentials, ensuring agents cannot bypass policy checks. It records every read and write, maintaining a journaled lineage graph. For developers, TrustGraph offers SDKs for Python and integrations with LangChain, OpenAI Agents SDK, and Claude Agent SDK. The system supports a 'needs approval' state for actions that require human sign-off, adding a layer of operational control without breaking automation.
Key Takeaways
- Provenance Tracking: TrustGraph tracks data lineage from source to output, ensuring restrictions follow the data through multi-step agent workflows.
- Deterministic Security: The decision engine is code-based and deterministic, avoiding the unpredictability of LLM-based guardrails.
- Gateway Enforcement: The architecture uses a credential-holding gateway to prevent agents from bypassing policy checks by accessing tools directly.
- Experimental Status: Currently in experimental preview, the tool requires Rust 1.90+ and manual policy definition, though agents can assist in drafting policies.
The Bottom Line
TrustGraph is a necessary evolution for agentic systems, moving beyond simple 'can I read this?' checks to 'can I share this?' enforcement. While it requires strict discipline in defining sources and actions, it offers the only current path to preventing silent data leaks in complex, multi-tool agent pipelines. For teams deploying agents in regulated environments, this deterministic approach to output provenance is non-negotiable.