browser
Pass
Audited by Gen Agent Trust Hub on Jul 30, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides documentation for a built-in Browser SDK. It establishes clear operational boundaries, instructing the agent to never automate sensitive human-only steps like CAPTCHAs or logins, and instead use a dedicated handoff mechanism.
- [DATA_EXPOSURE]: The manual correctly identifies and warns the agent about the privacy implications of using the 'chrome' backend, noting that identity markers such as cookies and storage are shared with the user's profile. This transparency helps ensure the agent does not inadvertently expose or rely on isolated session states where they do not exist.
- [COMMAND_EXECUTION]: The skill allows the agent to generate and execute asynchronous Python code to interact with the browser. While this is a powerful capability, it is limited to a closed SDK surface defined within the platform environment, and no external code execution or shell access patterns were found.
- [PROMPT_INJECTION]: Instructions within the skill emphasize accuracy and truthfulness, directing the agent to report only observed facts and to never fabricate information if a step is blocked. These are positive behavioral constraints that reinforce safety protocols.
- [INDIRECT_PROMPT_INJECTION]: As a tool designed to read web content, the skill possesses an inherent attack surface for indirect prompt injection via data ingested through
page.snapshot(). - Ingestion points: Web content is ingested into the agent context via
page.snapshot()(SKILL.md). - Boundary markers: The instructions advise the agent to state only observed facts, but there are no explicit delimiters or instructions to ignore embedded commands within the ingested text.
- Capability inventory: The agent can perform actions like
page.click(),page.fill(), andpage.goto(), which could be exploited by malicious web content. - Sanitization: The skill does not mention sanitization of the snapshot text.
- This surface is typical for browser-based skills and is managed through the agent's core instructions and safety layers rather than the skill itself.
Audit Metadata