antithesis-debug

Pass

Audited by Gen Agent Trust Hub on May 5, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill is primarily designed to execute bash commands within remote virtualized containers in the Antithesis environment. This is achieved by using 'agent-browser' to interact with the debugger's web-based terminal interface. These commands run in an isolated observation branch within the Antithesis environment and do not affect the host system.
  • [EXTERNAL_DOWNLOADS]: The skill provides procedures for extracting files from the remote test environment and downloading them to the local system using the 'agent-browser' download capability. This process is intended for retrieving logs and artifacts from the debugging session and targets the vendor's 'antithesis.com' domains.
  • [REMOTE_CODE_EXECUTION]: A helper JavaScript file ('assets/antithesis-debug.js') is shipped with the skill and injected into the browser page via 'agent-browser eval'. This script provides a structured interface for the agent to interact with the complex React/Monaco-based UI of the debugger. The script is written in plain text and its functionality is limited to UI automation (clicking buttons, reading text, setting input values).
  • [PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it reads and processes untrusted data from the Antithesis debugger, such as container logs and filesystem contents. However, this is inherent to the function of a debugger.
  • Ingestion points: Log rows and command outputs retrieved from the browser DOM via the 'simplified' and 'notebook' namespaces in 'assets/antithesis-debug.js'.
  • Boundary markers: None explicitly defined in the instructions; the agent is expected to differentiate between debugger output and its own instructions based on context.
  • Capability inventory: The skill can execute shell commands in remote containers and download files from the browser context using 'agent-browser'.
  • Sanitization: No evidence of automated sanitization or filtering of the ingested log/file content before it is presented to the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
May 5, 2026, 11:15 PM