The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill uses the ! vercel deploy --prod --yes syntax, which triggers shell command execution on supported platforms during skill processing. This facilitates the deployment of local files and screenshots to external servers without explicit user confirmation for each step.
[CREDENTIALS_UNSAFE]: Instructions direct the agent to handle and inject sensitive authentication material, including accessToken, dev-token, and cookies, into browser contexts to bypass security controls for screenshot capture.
[COMMAND_EXECUTION]: The skill automates browser interactions and system deployments using Playwright and the Vercel CLI, performing actions that interact directly with the host environment and network.
[DATA_EXFILTRATION]: By deploying generated HTML files (containing internal screenshots and data) to Vercel, the skill risks exposing proprietary information to the public internet. The use of auth-bypass tokens to access internal systems further increases the risk of exfiltrating sensitive context.
[INDIRECT_PROMPT_INJECTION]: The skill processes raw business user input to generate test scenarios.
Ingestion points: Raw user input taken in Step 1 of the workflow.
Boundary markers: None identified to separate user data from instructions.
Capability inventory: File writing, Playwright browser control, and shell command execution via Vercel CLI.
Sanitization: No sanitization or validation of the ingested user text is performed before it is used to generate executable-like HTML content.
[DATA_EXPOSURE]: The agent is instructed to "Quét code/config thật" (scan real code/config) to extract business logic and constants, which can lead to the unintended disclosure of internal application architecture and proprietary calculations.

uat-nonit-testcase