ai-archive
Fail
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: HIGHCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute several shell commands (e.g.,
git ls-files,mv,git mv) using the<name>variable, which is derived from the user-supplied$ARGUMENTS. In Step 6, the commandgit ls-files --error-unmatch "<output_root>/<name>"uses double quotes, which in most shell environments still allows for command substitution (e.g.,$(...)) or variable expansion. If a folder name on disk or a user-provided argument contains shell metacharacters, it could lead to arbitrary command execution within the agent's environment. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local filesystem and interpolates it into manifest files, creating a vulnerability surface for indirect prompt injection.
- Ingestion points: User input through
$ARGUMENTS(SKILL.md Step 2), folder names within theoutput_root(SKILL.md Step 2), and contents of the## Statusand## Descriptionsections in projectREADME.mdfiles (SKILL.md Step 3, 7, and 8). - Boundary markers: The skill includes logic to prevent path traversal (e.g., checking for
..in Step 2) and restricts operations to anoutput_root. However, it lacks instructions for the agent to sanitize or ignore potential AI instructions embedded within the folder names or README descriptions before they are processed and written to the manifest files. - Capability inventory: The agent is granted capabilities to execute shell commands (
mkdir,mv,git), read various local files (.ai-skills.toml,README.md,AGENTS.md), and perform file writes/edits on manifest files. - Sanitization: Beyond path validation, there is no content-level sanitization or escaping performed on the strings extracted from project files before they are transplanted into the
README.mdmanifests.
Recommendations
- AI detected serious security threats
Audit Metadata