meta-apply
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of patch files (.diff) and manifest data (manifest.jsonl) produced by other agents. This creates a vulnerability where malicious code or instructions hidden within a patch could attempt to bypass the 'jury' reviewer or exploit the agent performing the application.
- Ingestion points: Staged files in the
.aris/meta/pending/directory. - Boundary markers: No specific delimiters or 'ignore' instructions are documented for the ingestion of the raw diff content.
- Capability inventory: The skill has access to powerful tools including
Bash,Write,Edit, and the ability tospawn_agentfor external reviews. - Sanitization: The skill relies on a 'fresh landing review' (a separate LLM instance) and manual human authorization as its primary defense against malicious content.
- [COMMAND_EXECUTION]: The skill uses shell commands via the
Bashtool to resolve file paths, read manifest files, and execute a local Python script (provenance.py). Command arguments are constructed using variables (e.g.,$TARGET,$AUTHOR,$JURY_REVIEW_ID) sourced from the manifest files staged by other agents, which could lead to command injection if those files are maliciously crafted. - [DYNAMIC_EXECUTION]: The skill dynamically spawns a new reviewer agent (
spawn_agent) at runtime to judge the staged patches. While this is used as a security measure, it involves invoking a high-capability model based on parameters derived from the current task context.
Audit Metadata