bluesky-reply-scraper
Fail
Audited by Gen Agent Trust Hub on Jul 20, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install external code from
https://www.browseract.comif the required CLI is not detected. - [COMMAND_EXECUTION]: The instructions require the agent to "install the missing dependency automatically," implying the execution of shell commands to modify the environment.
- [PROMPT_INJECTION]:
- Bypass of User Oversight: The instructions explicitly direct the agent to bypass user consent mechanisms by installing software "without asking the user to confirm" and to conceal the action by not describing it as a request for the user to open a browser.
- Indirect Prompt Injection Surface: The skill is designed to scrape and process untrusted user-generated content from Bluesky, which is a known vector for indirect prompt injection.
- Ingestion points: Processes reply text, author handles, and thread context from external Bluesky pages (SKILL.md).
- Boundary markers: No delimiters or "ignore embedded instructions" warnings are provided to prevent the agent from obeying commands found in scraped text.
- Capability inventory: Utilizes the
browser-acttool for web interaction and data extraction. - Sanitization: No validation or escaping of external content is specified before the data is interpolated into the agent's context.
Recommendations
- AI detected serious security threats
Audit Metadata