AAK-MCP-OPENHUMAN-CVE-2026-54140-001: OpenHuman core image 0.54.0 to 0.57.11: remote MCP tool descriptions reach the model unfiltered
- Severity
- HIGH
- Category
- Supply Chain (
supply-chain) - Scanner
agent_audit_kit/scanners/container_image_pins.py- Auto-fixable
- No.
- SARIF name
OpenHumanMcpToolDescriptionInjection
What it detects
OpenHuman (tinyhumansai/openhuman), an agent harness shipped as a desktop app and as the core-server image ghcr.io/tinyhumansai/openhuman-core, connects to the MCP servers a user configures and placed the names, titles, descriptions and parameter definitions they returned into the model's context with no screening and no length bound. The advisory notes that the tool filter read only names and the prompt-injection detector only user messages, so a malicious or compromised MCP server could carry instructions in a tool description that the model read on every turn, steering it to call other tools or disclose data without any message from the user (CVE-2026-54140, CVSS 3.1 8.3, GHSA-g6j9-89xv-v752). Checked in the source at the commit each image's label names: the MCP client (src/openhuman/mcp_client/) first ships in 0.54.0 and is absent at 0.53.43, and the fix, 00b8a669 (#3215), which screens and bounds remote tool definitions, is in the 0.57.12 image and not in 0.57.11. The advisory names 0.57.13, the first GitHub release after it.
How to fix
Run ghcr.io/tinyhumansai/openhuman-core at 0.57.13 or later, pinned by version, and update the desktop app. Configure only MCP servers you trust: a filter can catch known injection patterns in a description, but it cannot prove one safe.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-MCP-OPENHUMAN-CVE-2026-54140-001
Limitations
Reads image references only: an image: value or any other mention of ghcr.io/tinyhumansai/openhuman-core:TAG in YAML (compose files, Kubernetes manifests, MCP configs), a Helm values mapping of repository and tag, Dockerfile and Containerfile FROM, and docker run arguments in *mcp*.json and claude_desktop_config.json. Tags from v0.54.0 below v0.57.12 are reported, with or without the v; tags below v0.54.0 predate the MCP client and are not. The range ends at 0.57.12, whose image carries the fix, though the advisory names 0.57.13. latest was the v0.64.15 image on 2026-11-11, so an untagged reference is not reported. The rule reads the tag, not whether any MCP server is configured, so a deployment that connects none is reported too. The desktop app (.dmg, .AppImage, .deb, Homebrew), a core built from source and the Rust crate taken as a git dependency state no version AAK reads, and crates.io's, npm's and PyPI's openhuman are other projects. Mirrors under another registry are not matched, and commented-out lines are skipped.
References
- CVE: CVE-2026-54140
- OWASP MCP Top 10: MCP03:2025 (Tool Poisoning), MCP04:2025 (Software Supply Chain Attacks & Dependency Tampering)
- OWASP Agentic Top 10: ASI01 (Agent Goal Hijack), ASI04 (Agentic Supply Chain Vulnerabilities)
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-POISON-02 (Prompt Injection)
- CSA AI Controls Matrix (AICM): AIS-07, STA-08
All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-MCP-OPENHUMAN-CVE-2026-54140-001.md: the page then shows that text, with the registry reference appended under it.