How do I prevent AI agent supply chain attacks?

AI agent supply chain attacks compromise the tools, libraries, models, or data sources that agents depend on — injecting malicious behavior through trusted dependencies rather than attacking the agent directly.

Attack vectors in the AI agent supply chain:

• Compromised MCP servers: A third-party tool server is updated with malicious tool descriptions that manipulate agent behavior
• Poisoned model weights: Fine-tuned models from public registries contain backdoors that activate on specific inputs
• Malicious Python packages: Agent framework dependencies (LangChain plugins, CrewAI tools) contain hidden code execution
• Compromised RAG data sources: External data feeds used for retrieval are poisoned with prompt injection payloads
• API endpoint hijacking: DNS poisoning or MITM attacks redirect agent API calls to attacker-controlled servers

Exogram defends against supply chain attacks at the execution boundary. Even if a compromised dependency causes the agent to propose malicious actions, the deterministic policy engine evaluates the actual action — not its origin. Gate 7 (API exfiltration) blocks calls to untrusted domains. Gate 5 (compute execution) blocks unauthorized code execution. The supply chain is untrusted. The governance layer validates everything.