Vulnerability  ·  2026-11-10

AshAi tool error messages leak internal module names and record data to MCP clients and the model (CVE-2026-119613)

VulnerabilityLow impactGlobalCVE-2026-119613
AshAi.ToToolError incorrectly returns the raw error field of Ash.Error.Unknown.UnknownError, and AshAi.Tool.Errors.format/1 forwards that text to MCP clients and to the model. The full Exception.format/2 or inspect/1 output can include whole records with fields the tool never intended to expose. CWE-209, CVSS 4.0 5.3 (Medium).
In an agentic/MCP setup, a tool that leaks internal module names and record field values to the LLM or to an MCP client can expose sensitive business data and aid subsequent prompt-injection or reconnaissance — and the errors were not logged, weakening incident response.
An authenticated caller makes an AshAi tool call fail in a way that cannot be JSON-encoded, causing the raw exception/inspect output (module names and inspected record values the tool never returns) to be sent to MCP clients (tool calls and resource reads) and to the model in AshAi.ToolLoop
ash_ai (Elixir Ash framework AI tools) 1.5.0 before 1.5.1
Upgrade to ash_ai 1.5.1, which logs errors under an ID and returns only that ID to the caller; constrain which data a failing tool can expose (commit af7987769d78ed070c8a9f6e6ee542d91744eae6 / advisory GHSA-vhmw-fwrr-r6c6)
NVD REST record for CVE-2026-119613Erlang/Elixir CNA recordash_ai GitHub advisory
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →