The security checklist in addyosmani/agent-skills: When agents write or review code, first map trust boundaries, list assets, run through STRIDE, and then implement protections for input validation, authentication sessions, security headers, SSRF, dependency supply chain, personal data, and AI features according to the three-tier rules of "Must Do / Ask a Human First / Never Do."
skills.sh install count 73493 (+959 / 24h)
Security Audit: Not audited
- Not yet scanned
Rationale
- The three-tier boundaries are clearly defined: Parameterized queries and using bcrypt / scrypt / argon2 for passwords are "Must Do"; adding new authentication flows, changing CORS, adding upload interfaces, or elevating permissions require prior human approval; committing secrets, logging passwords, or placing LLM output into eval or innerHTML are "Never Do."
- The dependency supply chain section requires disabling dependency install scripts before the first installation and allowing only the minimum necessary, while noting that audit tools can only detect known vulnerabilities and cannot identify new malicious packages or typosquatting.
- AI features are listed separately: Treat model output as untrusted input; prompt injection cannot be prevented solely by system prompts; validate tool parameters individually and require confirmation for destructive operations.
- skills.sh has 959 installs in the last 24 hours, with a cumulative total of 73493; the same repository contains another 24 skills (indicating only level of attention).
Usage
- Triggered when writing code that handles user input, authentication, uploads, or external calls, or when conducting security reviews; no account required.
- Run the package manager's built-in audit against the checklist before publishing.
Risks
- Not audited in this period, fixed at C-level, and install commands are not displayed; SKILL.md consists purely of processes and checklists, which will cause agents to run dependency audit commands.
- The checklist is biased towards web backends, with limited applicability to desktop or embedded projects; it is not an automated scanner, and conclusions still require manual review.