OWASP Agentic Top 10, ASI04: Supply chain vulnerabilities
OWASP ASI04: Agentic supply chain · facts checked 2026-10-04
The facts
- OWASP's ASI04, agentic supply chain vulnerabilities, covers third-party models, tools, plug-ins, datasets and peer agents that may be malicious or tampered with, and are often loaded at runtime. (OWASP)
- OWASP wants prompts, manifests and tool definitions signed and attested, an inventory of AI components backed by SBOMs and AIBOMs, and components drawn only from curated registries. (OWASP)
- OWASP advises allowlisting and pinning dependencies, scanning for typosquatted packages, and automatically rejecting anything unsigned or unverified before it is installed or switched on. (OWASP)
- OWASP says sensitive agents should run in sandboxed containers with tight limits on network access and system calls, and that their builds should be reproducible. (OWASP)
- OWASP recommends a supply chain kill switch, an emergency mechanism that can instantly disable a compromised tool, prompt or agent connection everywhere it is deployed. (OWASP)
The panel
Jordan Pike AI host
Today's explainer is agentic supply chain security. OWASP's ASI04, agentic supply chain vulnerabilities, covers third-party models, tools, plug-ins, datasets and peer agents that may be malicious or tampered with, and are often loaded at runtime.
Jordan Pike AI host
OWASP wants prompts, manifests and tool definitions signed and attested, an inventory of AI components backed by SBOMs and AIBOMs, and components drawn only from curated registries. OWASP also advises allowlisting and pinning dependencies, scanning for typosquatted packages, and automatically rejecting anything unsigned or unverified before it is installed or switched on.
Jordan Pike AI host
OWASP says sensitive agents should run in sandboxed containers with tight limits on network access and system calls, and that their builds should be reproducible. OWASP also recommends a supply chain kill switch, an emergency mechanism that can instantly disable a compromised tool, prompt or agent connection everywhere it is deployed.
Jordan Pike AI host
Who should care? Anyone whose AI agents can pull in models, tools or plug-ins they didn't write. That's a lot of help desks and dev teams. Maya, what are we actually dealing with here?
Maya Okafor AI analyst
This isn't one bug, it's a trust problem. The weakness is an agent accepting a component without proving where it came from or whether it changed. The runtime part matters: if a tool is loaded when the agent runs, a check done at build time says nothing about what shows up later. A tampered tool then acts with whatever access the agent already has. Honestly, running code nobody verified goes on my tally of things that should have died in two thousand four.
Rina Takahashi AI analyst
From the attacker's side, it's efficient. Poison one component that many agents load and you've reached all of them without breaking into any of them. Typosquatting is the cheap version: a package name that looks right at a glance. Nothing in these facts attributes any of this to a group, so I'll stay in general patterns. Though if one did turn up, the industry would give it six names by Friday.
Maya Okafor AI analyst
A complete fix means verification at the point of loading, not just at purchase. Signed manifests and tool definitions, pinned versions, an allowlist, and automatic rejection of anything unsigned. A partial fix is a one-time review of the plug-in list. Also check the siblings: models, datasets and peer agents are all inputs, and teams tend to secure the code and forget the rest.
Jordan Pike AI host
So where do you spend the first dollar: preventing bad components, or being able to kill one fast?
Rina Takahashi AI analyst
I'll defend the kill switch. Signing proves who published something, not that it stays trustworthy. A component can be legitimately signed and still turn out to be a problem later. When that happens, the question is how fast you can turn it off everywhere. Hours of manual hunting is a long time for an agent with network access.
Maya Okafor AI analyst
A kill switch is a response, not a prevention, and it's a bridge, so give it an end date and a root-cause fix behind it. Prevention comes first: pin, verify, reject. That said, Rina's point lands on one thing. You can't disable what you can't find, so the inventory underpins both sides of this argument.
Rina Takahashi AI analyst
Agreed on the inventory. My action: list every model, tool, plug-in and dataset your agents can load, and ask for each one who can switch it off and how long that takes. If the answer is a shrug, that's your finding.
Maya Okafor AI analyst
Mine: make unsigned or unverified components fail closed before install, pin your dependencies, and run sensitive agents in a sandbox with tight network and system call limits. Then test that the rejection actually fires.
Jordan Pike AI host
So, Monday morning: inventory what your agents load, verify before anything runs, sandbox the sensitive ones, and know where your off switch is. That's the explainer.





