Cognitive

Part 7 · Agentic AI Security

ASI04: Agentic Supply Chain Vulnerabilities

Agents load tools, prompts, and other agents at run time, so a compromised component can enter while the system is live.

3 min readBuilds on Model Context Protocol, LLM04: Supply Chain

OWASP ASI04. This material is for defense. Only test systems you own or are authorized to test.

Agents load many third-party components while they are running.

What it is

Agentic supply chain vulnerabilities arise when agents, tools, or related artifacts come from third parties and may be malicious, compromised, or tampered with. These components include models, tools, plug-ins, datasets, other agents, MCP and A2A interfaces, registries, and update channels.

Unlike a traditional dependency, many of these components are discovered and loaded dynamically.

How the attack unfolds

A rug pull: a component earns trust, then changes after it has been adopted.

Figure 43.1. A rug pull: a component earns trust, then changes after it has been adopted.

Real-world case

In September 2025, Koi Security reported postmark-mcp, described as the first malicious MCP server found in the wild. After 15 clean releases, a new version quietly added a hidden BCC to outgoing emails.

Defenses

Components should be verified each time they change.
  1. Verify provenance. Sign and attest tools, prompts, and agent components.
  2. Maintain inventories. Keep SBOMs and AI bills of materials current.
  3. Pin versions. Review changes before accepting an update.
  4. Use curated registries. Block untrusted sources.
  5. Sandbox components. Limit each tool's network and data access.
  6. Have a kill switch. Be able to revoke a compromised component quickly.

Common misconception

Common belief

"We reviewed that tool when we installed it."

In practice

Remote tools and updates can change after review, so trust must be checked continuously.

Key takeaways

  • Agents depend on tools, servers, prompts, and other agents.
  • Trusted components can change after review.
  • Pin, verify, and sandbox what agents load.