pr-screenshot-docs
Fail
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: HIGHDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill instructs the agent to use
curlto upload files tohttps://0x0.st, a public and ephemeral file-hosting service. This facilitates the transmission of local files to an external, non-whitelisted domain, which can be abused to exfiltrate sensitive project data or credentials beyond the intended screenshots.\n- [COMMAND_EXECUTION]: The inclusion of the shell commandcurl -F 'file=@screenshot.png' https://0x0.stprovides a functional template for network-based file transfers. This allows the agent to execute commands that move data from the local environment to third-party servers, bypassing organization-controlled storage solutions.\n- [EXTERNAL_DOWNLOADS]: The skill encourages the use of an untrusted third-party service (0x0.st) for hosting documentation assets. This introduces a dependency on unverified infrastructure that is not managed by a trusted vendor or well-known service provider.\n- [INDIRECT_PROMPT_INJECTION]: The skill uses browser tools to process and document UI elements, creating a surface for indirect instructions to enter the agent context.\n - Ingestion points:
browser_snapshotandbrowser_take_screenshot(SKILL.md).\n - Boundary markers: None; there are no instructions to the agent to treat UI content as untrusted or to ignore embedded text instructions.\n
- Capability inventory: The agent has shell access via
curland the ability to read/write files and interact with web content.\n - Sanitization: None; the skill does not specify any validation or filtering of the content captured from the browser.
Recommendations
- AI detected serious security threats
Audit Metadata