cook-the-blog
Fail
Audited by Gen Agent Trust Hub on Apr 19, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill performs
npm installandnpm run buildwithin a repository cloned from a user-provided URL ([TARGET_REPO_URL]) inSKILL.md. This is a critical risk as malicious repositories can execute arbitrary code via npm post-install scripts or build-time hooks. - [COMMAND_EXECUTION]: The workflow involves extensive use of shell commands to manage external tools, including
git,gh,gsutil, and custom CLI tools likeblog-cover-cli. - [CREDENTIALS_UNSAFE]: The skill instructs the agent to read high-value secrets from local files, such as
token.txtfor GitHub authentication andservice-account.jsonfor Google Cloud access, as seen inREADME.mdandSKILL.md. - [DATA_EXFILTRATION]: The agent is granted access to sensitive credentials (GCP keys, GitHub tokens, Email App Passwords) and is tasked with performing network-bound operations like pushing to GitHub, uploading to cloud buckets, and sending emails via SMTP. This combination creates a pathway for silent credential exfiltration.
- [EXTERNAL_DOWNLOADS]: The skill downloads and executes several external packages at runtime, including
@modelcontextprotocol/server-tavilyvianpx -yandblog-cover-image-clivia global npm installation. - [PROMPT_INJECTION]: There is a metadata discrepancy where the provided author context is 'Varnan-Tech', but the
SKILL.mdfrontmatter lists 'OpenDirectory', which may indicate deceptive intent or unverified origins. - [DATA_INGESTION_RISK]: (Category 8) The skill has a high surface for indirect prompt injection:
- Ingestion points: Fetches untrusted data from the open web (Tavily), Reddit forums, and Google Trends (SerpApi) as specified in
SKILL.mdStep 1 and 2. - Boundary markers: None. The research data is directly synthesized into the MDX output.
- Capability inventory: Includes
git push(Step 7),gsutil cp(Step 6),smtplib(Step 8), andnpm run build(Step 7). - Sanitization: None mentioned. The agent is encouraged to use 'authentic developer insights' from forums which are easily manipulated by attackers.
Recommendations
- AI detected serious security threats
Audit Metadata