Cognition AI has dropped a new repository of agent skills designed to automatically identify and patch critical security vulnerabilities in PHP, Node.js, and GitLab environments. Published on October 25, 2026, the 'devin-vuln-fixes' repo provides three specific SKILL.md files that work across a wide array of coding agents, including Devin, Claude Code, Codex, Cursor, and GitHub Copilot. The release targets urgent remote code execution (RCE) flaws, most notably CVE-2026-110264 in self-managed GitLab, which carries a CVSS score of 10.0.

The Vulnerability Landscape

The skills are pinned to advisories from October 21, 2026, addressing unauthenticated RCEs in every web SAPI for PHP (CVE-2026-109618 and CVE-2026-109619) and a Brotli-related RCE in Node.js (CVE-2026-111497). For GitLab, the fix targets the critical CVSS 10.0 flaw in Omnibus, Docker, Helm, and IaC deployments. The repository specifies exact fixed releases, such as PHP 8.5.12 and GitLab 19.5.1, ensuring that agents do not guess versions but instead verify that the specific artifact containing the fix exists before editing.

Agent-Neutral Execution Model

Unlike vendor-locked plugins, these skills are built to be portable, relying on the Agent Skills spec rather than proprietary tool definitions. The execution model is rigorous: the agent performs a read-only inventory of Dockerfiles, CI configs, and infrastructure code, then stops at an approval gate. Only after human approval does the agent patch the files, run baseline and post-patch tests inside the production image, and generate a single focused PR with rollback plans. This design prevents the 'hallucinated fix' problem common in AI-assisted coding, as the skills explicitly avoid guessed versions and require test validation within the actual runtime environment.

Installation and Integration

Developers can install the skills via npx or through plugin marketplaces for specific agents like Claude Code and Factory Droid. The CLI detects installed agents and allows for global or project-specific installation. For those preferring manual control, the skill folders can be copied directly into standard directories like .agents/skills/ or .claude/skills/. The system requires Node.js 22.20 or later and demands read access to repos, plus Docker and registry access for the build and test phases. If public hosts are blocked, the skills instruct the agent to mark targets as unverified, deferring proof to the CI pipeline.

Key Takeaways

  • The skills target critical CVSS 10.0 and unauthenticated RCEs in GitLab, PHP, and Node.js.
  • They are portable across 70+ coding agents by using the standard SKILL.md format.
  • The workflow includes a human approval gate and mandatory in-image testing to prevent regressions.
  • Version targets are pinned to October 21, 2026 advisories, with automatic updates via npx skills update.

The Bottom Line

Cognition is quietly winning the infrastructure wars by making AI agents actually safe to use for production patches, not just for generating boilerplate. This isn't just about convenience; it's about trust. By enforcing an approval gate and verifying artifacts exist, these skills turn AI from a risky autopilot into a competent junior engineer.