Supply chain and software flaws in AI tools
Malicious packages, poisoned skills and plugins, planted agent configuration, and ordinary vulnerabilities in MCP servers, coding agents and agent frameworks.
- LLM04:2026 Supply Chain
- ASI04:2026 Agentic Supply Chain Vulnerabilities
- ASI05:2026 Unexpected Code Execution
- AML.T0010 Supply Chain Compromise
- AML.T0109 AI Supply Chain Rug Pull
- EU AI Act 15(5) Model poisoning
How it works
- 1Agents run code from registries: MCP servers started with npx or uvx, skills, plugins and the frameworks themselves.
- 2A malicious package runs with the agent's permissions; a vulnerable one gives an attacker a way in.
- 3A repository can also carry agent configuration, such as a hook that runs a command when the folder is opened.
Real cases on the map
- CriticalMLflow: Unauthenticated full-read SSRF in webhook delivery: _validate_webhook_url bypassed via unvalidated HTTP redirects (and DNS rebindingmlflowExploited
- CriticalLiteLLM has SQL Injection in Proxy API key verificationlitellmExploited
- CriticalMarimo: Pre-Auth Remote Code Execution via Terminal WebSocket Authentication BypassmarimoExploited
- CriticalUnauthenticated Remote Code Execution in Langflow via Public Flow Build EndpointlangflowExploited
- Criticaln8n Vulnerable to Remote Code Execution via Expression Injectionn8n, n8nExploited
- CriticalLangflow CORS misconfiguration enables Account Takeover and RCElangflowExploited
Published advisories whose own titles describe this technique.
What to watch for
- A package version with a published advisory
- A package reported as malicious
- Hooks or MCP commands that download and run code
How to defend
- Pin and review the packages and servers agents run.
- Check configs and installed versions against advisories.
- Review agent configuration in repositories like code.