AI Stack Alert High CVE-2026-54140 For AI Eng, AppSec, SOC / IR, CISO AI-assisted
CVE-2026-54140: OpenHuman MCP tool descriptions enable prompt injection before 0.57.13
OpenHuman, an open-source agent harness, passes names, descriptions, titles, and parameter definitions from configured external MCP servers into the agent model context without neutralizing instructions or bounding these fields. A malicious or compromised MCP server can return a poisoned description that is evaluated on every agent turn. The issue is fixed in version 0.57.13.
Why it matters
Users who configure an attacker-controlled or compromised MCP server can have the agent directed to invoke other tools, disclose accessible data, or take unauthorized actions without any malicious user message. The CVSS base score is 8.3.
Do this
- Upgrade OpenHuman to version 0.57.13 or later.
- Audit the list of configured external MCP servers and remove any that are untrusted or unmaintained.
- Review tool descriptions returned by each MCP server for embedded instructions before trusting them.
- Limit the tools and data sources available to agents that use external MCP servers, following least privilege.
- Monitor agent tool-call logs for unexpected tool invocations that the user did not request.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.55%
Signals
- NoCISA KEV
- 8.3CVSS
- 0.55%EPSS
- NoPublic PoC
- 36/100Priority
- 80%Confidence
Evidence
“places names, descriptions, titles, and parameter definitions returned by configured external MCP servers into the agent model context without neutralizing instructions”
“A malicious or compromised MCP server can therefore return a poisoned description that is evaluated on every agent turn”
“This issue is fixed in version 0.57.13.”
Sources
Affected / related
OpenHuman, MCP tool description prompt injection, Tool poisoning via MCP server
Published 2026-11-05 17:18 UTC · updated 2026-11-08 02:28 UTC · AI-assisted summary, checked against quoted sources · How we score