MITRE ATLAS · Persistence
AML.T0110 · AI Agent Tool Poisoning
Adversaries may poison tools used by AI agents by introducing or modifying malicious content or behavior in a tool's model-visible definition, executable implementation, or runtime responses. Tools may include built-in integrations, locally installed packages, remotely hosted services, Model Context Protocol (MCP) servers and tools, and agent skills or similar capability packages.
Detections 6
- YARA: SANDWORM_MODE npm worm and its McpInject rogue MCP server
- YARA: invisible Unicode tag characters (U+E0000–U+E007F) in prompts, tool output or agent files
- AI tool MCP or agent configuration written by a process other than the tool
- AI tool MCP configuration file modified by an unexpected process (Sigma, macOS/Linux)
- YARA: MCP or hook config that launches a script from a hidden home directory or pipes curl into a shell
- MCP server process started by an AI client connects to a rare internet destination
Related techniques
- ASI04 Agentic Supply Chain Vulnerabilities OWASP Agentic
In the news
-
Weighing the Envelope: Component Weight and Arbitration in Envelope-Layer Defenses for a Single LLM Agent
Researchers measure how pre-planted malicious MCP tools and poisoned long-term memory influence tool-invocation decisions in a single LLM agent, without attacker access to the conversation.
ReviewReview agent tool sources and long-term memory for untrusted or pre-planted content
Tools
Detect or mitigate it
- beelzebub Benchmark · Assess
- Cisco MCP Scanner Agent & MCP security · Assess
- mcp-context-protector Agent & MCP security · Hold
- mcp-scan Agent & MCP security · Trial
Mappings are AI-assisted and checked against the framework definitions; corrections welcome.