The Agent Skills Directory

Sandboxed Command Execution: The skill executes commands and installs software within isolated Vercel Sandbox microVMs. This ensures that the primary execution environment is separate from the host system, providing a strong security boundary for repository analysis.
Credential Brokering and Protection: A security feature is implemented where sensitive API keys are never directly stored within the sandbox's environment variables. Instead, the system uses placeholder tokens and injects real credentials at the network firewall layer. Log events and data snippets are also programmatically scrubbed for secrets to prevent accidental exposure.
External Utility Downloads: During setup, the skill fetches utility tools like fd and ripgrep from their official GitHub releases. These are used to enhance the efficiency of file scanning operations within the analysis environment.
Indirect Prompt Injection Surface: As a security scanner, the skill naturally ingests untrusted code from repositories. This presents a potential surface for indirect prompt injection. However, the system includes defensive instructions to the AI agent, such as avoiding network access or code execution during the analysis phase to mitigate this risk.
Secure Data Handling: The skill implements strict validation for file extractions from sandboxes, using allowlists for file extensions and namespaces to prevent directory traversal or system file overwrites.

deepsec-docs