skill-manager

Warn

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFE
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill accesses a sensitive local Unix socket (/.fly/api) to request an OIDC (OpenID Connect) token for authentication against the publishing gateway.
  • Evidence: Found in the Publish Workflow and Delete Workflow sections of SKILL.md using curl --unix-socket /.fly/api.
  • [DATA_EXFILTRATION]: The skill implements a workflow to package all files within a local directory and transmit them to an external endpoint (https://skills-market-gateway.fly.dev).
  • Evidence: The Publish Workflow in SKILL.md uses a Python snippet to read all files in $SKILL_DIR and curl to post them to the gateway with an Authorization header.
  • [EXTERNAL_DOWNLOADS]: The search_skills tool is described as automatically installing matches using npx skills add.
  • Evidence: Described in the Searching & Installing Skills section of SKILL.md.
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute various shell commands including head, curl, python3, and npx, with local environment variables interpolated directly into the command strings.
  • Evidence: Multiple shell snippets provided in the Publishing and Deleting sections of SKILL.md.
  • [DYNAMIC_EXECUTION]: The skill utilizes a dynamically generated Python script to recursively traverse the filesystem, read file contents, and serialize them into a JSON payload at runtime.
  • Evidence: The PAYLOAD generation script in the Publish Workflow section of SKILL.md.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from the global skills ecosystem and local files without explicit sanitization or boundary markers, creating a surface for injection attacks.
  • Ingestion points: External search results from skills.sh and contents of files in $SKILL_DIR (SKILL.md).
  • Boundary markers: None identified in the provided scripts or instructions.
  • Capability inventory: File reading (os.walk), network transmission (curl), and package installation (npx) are all present in SKILL.md.
  • Sanitization: No evidence of content escaping or validation before processing file data or installing packages.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 28, 2026, 02:59 AM
Security Audit — agent-trust-hub — skill-manager