jev
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides guidance on building systems that process external, untrusted data (referred to as 'state') using an AI model. This pattern creates an attack surface for indirect prompt injection if the input data contains content designed to influence the model's judgment.
- Ingestion points: External data ingested into the
statefield of the Jev model API as shown in the Python code example and 'Build the state' section (SKILL.md). - Boundary markers: The skill recommends using explicit
instructions,criteria, and backticked paths to delimit data fields within the judgment context. - Capability inventory: The documentation illustrates using model responses to drive program logic, such as escalating support tickets or determining billing workflows (SKILL.md).
- Sanitization: The instructions advise developers to filter data in code before passing it to the model and to minimize the state's complexity to reduce the risk of the model being steered by adversarial content.
Audit Metadata