WeaknessASC-2026-0004Major weakness
Agentic supply chain compromise
CriticalMitigated
Malicious or tampered models, MCP servers, tool descriptions, prompt packs or plugins enter the agent at runtime.
How to detect it
- Model files in pickle-based formats
- MCP server not pinned by version or digest
- Tool descriptions changed between versions (rug pull)
- Package published recently by a new maintainer (typosquat signals)
How to resolve it
- Curate every model, MCP server and plugin through AI Security Curator before use
- Pin by version or digest and verify signatures or hashes
- Prefer safetensors for model weights
- Diff tool lists on every upgrade and re-approve changes
- Keep an AI bill of materials (AIBOM)
How to investigate
- Compare the installed version and hash against the approved one
- Diff the current MCP tool list against the approved fingerprint
- Check for publisher or maintainer changes and unexpected network egress