atr-browser

Fail

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: HIGHDATA_EXFILTRATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
  • [DATA_EXFILTRATION]: The skill provides tools that can expose sensitive browser data to the agent's context or local system.
  • The eval command is explicitly documented in references/commands-reference.md as a way to retrieve authentication tokens using window.localStorage.getItem('token').
  • Inspection commands like html, text, snapshot, and ask return page content which may contain PII or session-specific information.
  • Commands like screenshot --file and download-images save data to the /tmp/ directory, which could be accessible by other local processes.
  • [DYNAMIC_EXECUTION]: The atr browser eval <script> command enables the execution of arbitrary JavaScript within the browser context.
  • This allows for complex, runtime-generated logic that can bypass standard interaction patterns and directly manipulate the DOM, cookies, or browser APIs.
  • [INDIRECT_PROMPT_INJECTION]: The skill presents a large attack surface for indirect prompt injection because it processes untrusted data from the web and possesses high-impact interaction capabilities.
  • Ingestion points: Untrusted data enters the agent context through snapshot, html, text, ask, and clean-snapshot commands.
  • Capability inventory: The skill allows the agent to click, fill forms (potentially with credentials), navigate to new URLs, and eval JavaScript.
  • Boundary markers: The instructions lack explicit delimiters or warnings to treat ingested web content as untrusted data, increasing the risk that the agent will follow instructions embedded in web pages.
  • Sanitization: There is no evidence of sanitization or filtering of external content before it is processed by the agent or the sub-agent used by the ask command.
  • [COMMAND_EXECUTION]: The skill uses the Bash(atr browser:*) tool to manage the browser lifecycle and perform automation.
  • While the tool is scoped, the broad range of subcommands allows for deep control over the browser environment, including network inspection and console access.
  • [CREDENTIALS_UNSAFE]: The documentation encourages the retrieval of secrets from the browser environment.
  • Example usage in references/commands-reference.md shows how to use eval to access sensitive authentication tokens, which may lead to insecure handling of these credentials in the LLM context.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Nov 12, 2026, 08:49 AM