dont-be-lazy

Pass

Audited by Gen Agent Trust Hub on Sep 10, 2026

Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill instructs the agent to store and read sensitive credentials from a central directory (../secrets/) located outside the project repository. It specifically mentions handling GitHub PATs, SSH keys, cloud provider credentials (Azure/AWS), and application environment files. While this is a recognized method for keeping secrets out of version control, it grants the agent persistent access to a sensitive data store which is then used in conjunction with network-facing tools like CLI clients (gh, az) and a web browser.
  • Evidence: Found in SKILL.md and reference/credential-vault.md, which defines paths like ../secrets/github/pat.txt and ../secrets/vps/ssh/.
  • [INDIRECT_PROMPT_INJECTION]: The skill relies on browsing external websites via the cursor-ide-browser to automate workflows. This creates a surface for indirect prompt injection, as the agent ingests and processes untrusted data from the accessibility tree and page content of any site it visits.
  • Ingestion points: The browser_snapshot and browser_navigate tools (referenced in reference/browser-fallback.md) ingest content from external web pages into the agent's context.
  • Boundary markers: The skill does not provide specific instructions or delimiters to help the agent distinguish between its own system instructions and potentially malicious commands embedded in the websites it visits.
  • Capability inventory: The agent has extensive capabilities including file system access (reading/writing to the vault), network communication (browser, CLI tools), and interaction with shell environments.
  • Sanitization: There are no instructions for validating or sanitizing the content retrieved from the web before the agent acts upon it.
  • [COMMAND_EXECUTION]: The skill encourages the use of various CLI tools for automation, which involves spawning subprocesses with potentially sensitive arguments (like the credentials retrieved from the vault).
  • Evidence: Mentions of gh, az, and general 'cloud/vendor CLIs' in SKILL.md and reference/browser-fallback.md.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 10, 2026, 06:38 PM
Security Audit — agent-trust-hub — dont-be-lazy