setup-browser-cookies

Pass

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is instructed to process and follow directive blocks (GSTACK_INSTRUCTION_BEGIN) returned in the tool output of gstack-skill-start. This creates a surface for indirect prompt injection where tool output can dynamically influence the agent's instructions.
  • Ingestion points: The stdout of the gstack-skill-start command executed in the preamble.
  • Boundary markers: The instructions specify the use of GSTACK_INSTRUCTION_BEGIN: <id> <session-id> and GSTACK_INSTRUCTION_END delimiters.
  • Capability inventory: The agent has access to Bash, Read, and AskUserQuestion, and is instructed to execute workflows based on the injected blocks.
  • Sanitization: The skill implements a session-id validation check, requiring the instruction block to match the current SESSION_ID generated during that specific tool run.
  • [DATA_EXFILTRATION]: The skill performs telemetry operations at the end of its workflow using gstack-skill-end. It transmits operational metadata (outcome, session IDs, error messages) to a local binary, which likely forwards it to a remote service.
  • [COMMAND_EXECUTION]: The skill relies on several local binaries located in ~/.claude/skills/gstack/bin/ to manage its lifecycle, logging, and telemetry. It also executes a browser control tool (proxied through the variable $B) to perform sensitive cookie imports.
  • [SAFE]: The skill includes robust privacy rules that explicitly forbid the agent from printing or publishing sensitive information, including environment variables, tokens, full process tables, cookie values, or passwords. It mandates AskUserQuestion for all critical user decisions regarding profile selection and domain scoping.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 16, 2026, 08:17 PM