Devin agent skills that patch the GitLab CVSS 10.0, PHP and Node.js RCEs

Cognition launched a suite of three autonomous Agent Skills designed to detect and remediate severe security flaws across multi-repository enterprise environments. Published in github.com's report, the open-source release supplies automated repair routines for critical flaws including an unauthenticated remote code execution vulnerability rated at CVSS 10.0 in self-managed GitLab instances, multiple unauthenticated PHP2Shell vulnerabilities affecting every web Server Application Programming Interface, and the Node.js security update from October 21, 2026. The packages operate within any coding assistant compliant with the Agent Skills specification, spanning Devin, Claude Code, Codex, Cursor, Factory Droid, GitHub Copilot, Gemini CLI, OpenCode, Windsurf, and more than 70 additional developer tools.

This architectural guide examines the engineering mechanics behind the automated remediations, the structural safety safeguards enforced before code changes touch production branches, and the deployment workflows required for platform engineers and security teams. System administrators managing self-hosted infrastructure, fleet maintainers overseeing polyglot codebases, and developers integrating agentic workflows into continuous integration pipelines can evaluate how these skills isolate changes and prevent regressions. By codifying inventory discovery, artifact confirmation, containerized validation, and pull request generation into standardized specifications, the toolset provides an automated mechanism to eliminate high-severity attack vectors without destabilizing operational builds.

What Devin agent skills that patch the GitLab CVSS shipped

The release delivers three specialized routines packaged as SKILL.md configurations to address critical infrastructure vulnerabilities. The fix-gitlab-vuln module targets CVE-2026-110264, an unauthenticated remote code execution defect evaluated at the maximum severity rating of CVSS 10.0 that compromises self-managed GitLab deployments. This automation inspects and updates configurations across Omnibus packaging, standard Docker containers, Kubernetes Helm charts, and infrastructure-as-code manifests including Terraform, Ansible, Chef, and Puppet. Supported upgrade paths migrate vulnerable installations directly to secured patch levels across seven supported release series, specifically releases 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10, and 18.11.13.

Alongside the GitLab component, the repository packages fix-php-vuln and fix-node-vuln to resolve remote runtime execution threats in application stacks. The PHP module neutralizes PHP2Shell vulnerabilities tracked under CVE-2026-109618 and CVE-2026-109619 across every web SAPI by shifting environments to safe releases 8.2.35, 8.3.36, 8.4.27, or 8.5.12. Meanwhile, the Node.js module handles the October 21, 2026 security release containing the CVE-2026-111497 Brotli remote code execution bug, upgrading dependencies to versions 22.23.4, 24.22.1, or 26.12.1 while identifying and flagging end-of-life Node 20 environments.

What changed for builders in Devin agent skills that patch the GitLab

Rather than allowing autonomous assistants to blindly rewrite dependency files, the system enforces a deterministic, five-phase remediation cycle across every targeted repository. The agent first initiates a read-only inventory scan, parsing Dockerfiles, CI matrices, version locks, and configuration management manifests. It constructs a structured assessment detailing the target file, current runtime version, pin style, execution context such as production or continuous integration, existing test coverage, and confidence score. The operation halts at an explicit approval gate where engineers select which repositories to modify, ensuring no code edits happen without human sign-off.

After receiving approval, the agent confirms that the target release tag, package version, or cryptographic digest genuinely exists in the designated registry before writing edits, avoiding fabricated versions. The assistant runs test suites before and after the modification inside the newly constructed production container image rather than on the host machine. Any newly introduced failures or fresh deprecation warnings are classified as unacceptable regressions. Upon passing validation, the agent opens a single isolated pull request per repository containing exact before-and-after digests, automated test output, risk classifications, deployment plans, and rollback targets.

How to install or upgrade Devin agent skills that patch the GitLab CVSS

Engineers can provision the routines across development environments using the global package runner or dedicated marketplace integrations. Using Node.js 22.20 or newer, developers can execute the standardized installation command to configure all supported agents detected locally. The tool enables selective targeting of specific runtimes or the installation of individual skills through command-line flags. Future advisory updates are pulled down using the corresponding update command.

npx skills add CognitionAI/devin-vuln-fixes
npx skills add CognitionAI/devin-vuln-fixes -a claude-code -a codex -a cursor -a droid -a devin
npx skills add CognitionAI/devin-vuln-fixes --skill fix-php-vuln
npx skills update

Specialized environments offer native plugin registrations. Claude Code users can register the repository marketplace and install the package directly, while Codex and Factory Droid provide marketplace registrations that map into internal skill directories. Devin users on cloud workspaces can add the skills via the plugin marketplace or specify the repository within the Customize page under connected settings, whereas Devin CLI and Desktop users invoke the standard package installer.

/plugin marketplace add CognitionAI/devin-vuln-fixes
/plugin install vuln-fixes@cognition-skills

droid plugin marketplace add https://github.com/CognitionAI/devin-vuln-fixes
droid plugin install vuln-fixes@cognition-skills

git clone https://github.com/CognitionAI/devin-vuln-fixes
cp -r devin-vuln-fixes/skills/fix-php-vuln .agents/skills/

Gotchas and compatibility in Devin agent skills that patch the GitLab

Execution requires repository read access alongside local Docker daemon availability and access to the container image registries or internal package mirrors used by production continuous integration systems. When air-gapped systems or enterprise firewalls block outbound connections to public registries, the skills instruct the assistant to flag the selected target as unverified, delegating validation to isolated continuous integration pipelines. Engineers operating in restricted networks must ensure private mirror access is exposed to the local container environment to execute pre-commit image tests.

To preserve compatibility across dozens of disparate agent platforms, Cognition intentionally omitted the allowed-tools configuration field from the YAML frontmatter of the skills. Specification standards diverge across implementations: the base Agent Skills specification mandates a space-separated string, whereas Devin CLI expects a structured YAML list. Omitting this field guarantees portability across Claude Code, Cursor, Codex, Windsurf, and Devin, though organizations requiring strict tool whitelisting must fork the repository to add platform-specific configuration fields manually.

What to watch after Devin agent skills that patch the GitLab CVSS

Post-implementation tracking depends on version lock validation and advisory monitoring. Fixed versions within the skills are pinned to the security advisories published on October 21, 2026. The skills-lock.json manifest stores the exact repository hash active in the workspace, ensuring that patch definitions remain immutable until explicitly refreshed. Teams should incorporate routine updates into automated maintenance rotations to catch upstream revisions as vendor ecosystems publish subsequent stability patches.

Operational teams reviewing the generated pull requests must verify rollback targets before merging container upgrades into production branches. Because the agent creates dedicated branches carrying full before-and-after cryptographic image digests, deployment pipelines should validate these signatures against internal security scanners. Infrastructure platforms utilizing subagent architectures can spawn one independent worker per repository to process fleet-wide vulnerability campaigns simultaneously, whereas standalone agents execute the patching workflow sequentially across designated projects.