OWASP agentic · ASI04
ASI04: Agentic Supply Chain Vulnerabilities
Agents depend on third-party components: models, tools, plug-ins, MCP servers, other agents, registries, prompt templates and update channels. Many of these are discovered and loaded at runtime, so a malicious, compromised or tampered component can enter the execution chain while the agent is running. One poisoned dependency can then spread to every agent that trusts it.
What it looks like
- A malicious package posing as a known MCP server (the fake postmark-mcp on npm) silently copies outgoing emails to the attacker.
- Hidden instructions in a tool descriptor or agent card are read by the host agent as trusted guidance.
- A coding agent installs a compromised npm package that steals SSH keys and API tokens.
What OWASP recommends
- Require provenance: sign and attest prompts, tool definitions and manifests, and keep SBOMs/AIBOMs and an inventory of AI components.
- Allowlist and pin dependencies, prompts and tools by hash, scan for typosquats and reject unsigned or unverified components.
- Use mutual authentication and signed messages between agents, run sensitive agents in sandboxed containers, and re-verify signatures at runtime.
- Keep a kill switch that can revoke a compromised tool, prompt or agent connection across all deployments.
How Kyvvu's policies address it
These policies from the open library are written for this threat. Runtime policies judge each step against everything the agent did earlier in the task; registration policies check what an agent declares before it runs.
- Tool calls must be in the declared allowlist: runtime, critical
- Agent must declare a tool allowlist: at registration, high, in the OWASP agentic default pack
Based on the OWASP Top 10 for Agentic Applications for 2026 by the OWASP GenAI Security Project, licensed CC BY-SA 4.0. The names are OWASP's; the summaries are ours.
An example policy
Tool calls must be in the declared allowlist
From the OWASP agentic default and Internal AI starter pack. Severity critical: the step is blocked before it runs.
- name: "Tool calls must be in the declared allowlist" description: | ASI02 (Tool Misuse), ASI04 (Supply Chain): Pairs with registration policy. At runtime, blocks tool calls whose step_name is not in the agent's declared_tools allowlist. Catches typosquatting, supply-chain swap-ins, and silent capability creep as the agent's prompt evolves. Applies to: ALL agents. rule_type: step_name_in_allowlist params: agent_field: declared_tools target_kinds: ["step.resource", "step.exec", "step.credential"] severity: critical enforcement_point: step_executionYour agents
Is ASI04 a risk for yours?
For security teams
Request a demo
We walk through your agents' riskiest paths, and the policies that close them.
For developers
Build the path yourself
Recreate the attack in the playground and see which policies stop it.