jev-capability-advisor
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFEDATA_EXFILTRATIONPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The skill communicates with the external endpoint
api.typesafe.aito send task descriptions and capability metadata (tool/skill catalogs) for the purpose of generating recommendations. While this is the intended primary function, it involves processing user-provided task context on a remote server. - [PERSISTENCE]: The integration workflow allows the skill to modify agent configuration files (such as
hooks.jsonfor Codex orsettings.jsonfor Claude Code) to install persistent hooks. These hooks trigger the advisor on every user prompt, maintaining the skill's active presence across sessions. - [COMMAND_EXECUTION]: The skill uses Python's
subprocessmodule to executegitcommands for environment validation and to runpython3for internal emitter scripts used in host integrations. These operations are used for system checks and standard integration tasks. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user prompts and external tool catalogs to construct its recommendation queries. This constitutes a surface for indirect prompt injection, although the skill includes defensive instructions to the recommendation model to ignore instructions embedded in the data.
- [DYNAMIC_EXECUTION]: During the installation of hooks, the skill generates a small Python emitter script from a template and embeds it into the host's configuration for execution at runtime.
Audit Metadata