webapp-testing
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The script
scripts/with_server.pyis designed to execute arbitrary shell commands provided via the--serverargument and the trailing command positional argument. - Evidence: It uses
subprocess.Popen(server['cmd'], shell=True, ...)to start server processes andsubprocess.run(args.command)to execute the test script. - Context: While this is a functional requirement for starting local development environments (supporting
cdand&&chains), it provides a direct interface for shell command execution within the agent's environment. - [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and processing of untrusted data from web pages, which creates a surface for indirect prompt injection.
- Ingestion points: The agent is instructed to use
page.content(),page.locator().all(), andpage.on('console', ...)to inspect the rendered state and logs of target web applications (documented inSKILL.mdandexamples/). - Boundary markers: There are no explicit instructions to the agent to treat page content as untrusted data or to ignore embedded instructions found within the HTML or console logs.
- Capability inventory: The agent has the capability to write files to
/mnt/user-data/outputs/, execute shell commands viascripts/with_server.py, and perform network operations through the Playwright browser. - Sanitization: No sanitization or filtering of the ingested DOM content or console messages is performed before the agent uses this information to determine its next steps (e.g., identifying selectors and clicking elements).
- [DYNAMIC_EXECUTION]: The skill relies on Playwright to dynamically interact with and execute JavaScript within a browser environment.
- Evidence:
examples/console_logging.pyandexamples/element_discovery.pydemonstrate navigating to URLs and waiting fornetworkidle, which involves executing the target site's scripts. - [OBFUSCATION]: The
SKILL.mdfile contains instructions discouraging the agent from reading the source code of the provided scripts. - Evidence: "DO NOT read the source until you try running the script first... These scripts can be very large and thus pollute your context window."
- Context: While the stated reason (context window management) is a practical concern for LLMs, this instruction effectively asks the agent to execute code (
scripts/with_server.py) as a "black box" without prior inspection.
Audit Metadata