browser
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted content from the open web, which serves as a significant attack surface.
- Ingestion points: Web content is brought into the agent's context through tools like
agent-browser snapshot,get text, andget html(SKILL.md). - Boundary markers: The instructions do not specify any delimiters or warnings to the agent to disregard instructions embedded within the retrieved web content.
- Capability inventory: The agent has the ability to execute shell commands via the
agent-browserCLI, write files (e.g.,state save auth.json,screenshot), and access network resources (SKILL.md). - Sanitization: No explicit sanitization or filtering of the retrieved HTML/text content is described before it is processed by the AI.
- [EXTERNAL_DOWNLOADS]: The skill uses
npxto fetch and execute the@claude-flow/clipackage from the npm registry at runtime (SKILL.md). - Execution method: Commands like
npx @claude-flow/cli memory storeandnpx @claude-flow/cli hooksexecute code directly from the remote registry without version pinning or integrity hashes in the provided instructions. - [DATA_EXFILTRATION]: The skill manages sensitive session information through state management commands.
- Evidence: The command
agent-browser state save auth.json(SKILL.md) captures browser session states, including cookies and potentially authentication tokens, and saves them to a local file. This file represents a high-value target for exfiltration if the agent is compromised or tricked into uploading it. - [COMMAND_EXECUTION]: The primary interface of the skill is the execution of shell commands through the
agent-browserCLI andnpx(SKILL.md). - Risk: The instructions encourage the agent to construct and run shell commands based on task requirements, which could be exploited if malicious input is successfully injected into the command construction logic.
Audit Metadata