claude-agent-sdk
Fail
Audited by Gen Agent Trust Hub on Mar 16, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: Example code in the documentation demonstrates using the unsafe eval() function to process user-provided strings within a custom tool, which is highly susceptible to arbitrary code execution (File: references/tools-reference.md).
- [COMMAND_EXECUTION]: The skill explicitly provides and documents built-in tools (Bash, Python) for executing arbitrary shell commands and Python code, granting extensive control over the underlying system (Files: references/api-reference.md, references/tools-reference.md).
- [EXTERNAL_DOWNLOADS]: The documentation encourages the installation of external packages from PyPI (claude-agent-sdk) and NPM (@anthropic-ai/claude-agent-sdk). While the latter is a well-known organization, the former's provenance should be verified against the author's identity (File: SKILL.md).
- [DATA_EXFILTRATION]: Network-enabled tools such as Fetch and WebSearch are included, which, when combined with file access tools like Read, facilitate the potential exfiltration of sensitive data to external servers (File: references/tools-reference.md).
- [PROMPT_INJECTION]: The skill is designed to ingest untrusted data from the local file system and the web (Read, Fetch) and uses this data as context for an agent with high-privilege capabilities (Bash, Write). There are no documented sanitization procedures or boundary markers (Files: SKILL.md, references/tools-reference.md). * Ingestion points: Read, Grep, Fetch, WebSearch. * Boundary markers: None documented. * Capability inventory: Bash, Python, Write, Edit. * Sanitization: None documented.
Recommendations
- AI detected serious security threats
Audit Metadata