guarded-webfetch-codex
Fail
Audited by Gen Agent Trust Hub on Oct 24, 2026
Risk Level: HIGHDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [DATA_EXPOSURE]: The skill accesses sensitive configuration files from the user's home directory. Specifically,
scripts/quarantine-fetch-codex.shcopies~/.codex/auth.json(Codex CLI credentials) into a temporary isolated environment. While this is intended to allow the child process to authenticate with the Codex service, it involves the handling of sensitive credential files. - [REMOTE_CODE_EXECUTION]: The skill relies on the
codex exectool to run logic in a child process. It constructs shell commands that execute local Node.js scripts (http-fetch-codex.ts,pipe-sanitize-codex.ts) within this environment. The child process is configured with--sandbox "danger-full-access"to allow network requests, which is a high-privilege execution mode. - [COMMAND_EXECUTION]: Multiple scripts involve spawning shell processes and piping data between tools like
bash,node,codex, andjqto fetch, sanitize, and process external data. - [INDIRECT_PROMPT_INJECTION]: As a web fetcher, the skill is exposed to untrusted data from external URLs. It manages this risk through a comprehensive sanitization pipeline in
scripts/sanitize.tsthat filters out Unicode tags, zero-width characters, and known prompt injection markers (e.g., "ignore previous instructions"). It also uses an automated gating mechanism to prevent the child LLM from reading text if suspicious patterns are detected. - [DYNAMIC_EXECUTION]: The skill uses
node -ewithinscripts/quarantine-fetch-codex.shto dynamically parse JSON and extract data from the fetcher's output at runtime. - [OBFUSCATION]: Automated scanner flags for Unicode steganography are false positives. The scripts
scripts/sanitize.tsand others contain regex literals (e.g.,/[\u200B-\u200F\u2060\uFEFF]/g) used explicitly to detect and remove hidden characters from fetched content, not to hide malicious intent.
Recommendations
- HIGH: Downloads and executes remote code from: unknown - DO NOT USE without thorough review
Audit Metadata