CVE-2026-77956

AshAi dynamic prompt text could execute Elixir code

ash-project · ash_ai

First documented disclosure · Review checkpoint

Summary

Request text incorporated into a function-generated prompt could run as a template.

Severity and exploitation evidence

High · 8.9/10 · CVSS 4.0. Severity source.

Exploitation not established: Malicious exploitation is not established by the reviewed originals. Public demonstrations and catalog absence do not establish exploitation status.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The prompt action passed function-supplied text to EEx evaluation before requesting a model response.

Conditions required

A reachable prompt action builds its string from attacker-controlled request data. Static configured templates and ReqLLM.Context results are different paths.

Potential impact

Elixir code can execute with the service's permissions.

Affected versions and configurations

ash_ai: 0.1.0 through versions before 1.0.0.

Fix and mitigation status

1.0.0.

Practical next steps

Apply the issue-specific correction: 1.0.0.

Operating lesson

Keep dynamic prompt content separate from executable template source.

Evidence and limits

Earliest reviewed public evidence 2026-08-30; CVE publication 2026-08-31. CVSS 4.0 8.9 is attributed to EEF in the CNA record. The CNA describes remote unauthenticated access but its selected CVSS 4.0 vector is local with attack requirements present; that source discrepancy is preserved rather than rescored.

Sources