Latest in AI
Cognition Vulnerability Fixes Plugin: How It Works Across Coding Agents
The open-source skills package turns runtime patching into a reviewable workflow, but the agent's permissions still matter.
Cognition released Vulnerability Fixes on October 22, an open-source skills package for coding agents that helps inventory runtimes, prepare updates and open reviewable pull requests. It works with Devin and other compatible agents. The useful part is the sequence: inspect first, get approval, then change and test.
The official repository contains three skills covering PHP, Node.js and self-managed GitLab workflows, under an MIT license. This is a package of instructions for an existing coding agent, rather than a new model or a standalone scanner.
How the plugin works
Cognition's launch walkthrough describes a read-only inventory across repositories the agent can access. The workflow looks beyond application source files, including container definitions, CI configuration and infrastructure code.
It then asks which repositories to change. Higher-risk work, such as a major runtime upgrade or a base operating-system change, is supposed to be identified before editing starts. The workflow checks proposed versions against the relevant source, runs tests before and after the update, and prepares one pull request per repository.
That sequence is useful because a runtime update is rarely just a version-string edit. A team needs to know where the runtime is selected, which artifact gets built and what evidence will support the change. A pull request with that context is easier to review than a scattered collection of automated bumps.
Which agents can use it
The repository provides installation paths for Devin, Claude Code, Codex and Factory Droid, alongside the portable Agent Skills format used by other coding tools. The general installer is npx skills add CognitionAI/devin-vuln-fixes; its documented prerequisite is Node.js 22.20 or later.
In Devin cloud, Vulnerability Fixes is available through the plugin marketplace or the Customize page's repository option. Developers can also install individual skills instead of the entire package. The repository documents updating installed skills with npx skills update.
Open-source availability does not make the underlying agent's execution free. Model usage, build infrastructure and the coding product's own charges are separate considerations.
Approval instructions are not a permission boundary
The skill tells the agent when to ask; it does not replace the agent's permission system. Cognition warns that skills run with the permissions of the agent hosting them, and local products may not provide Devin's isolation.
The Node.js skill's implementation makes the intended controls concrete. Discovery must remain read-only. It requires approval before editing a repository and separate approval for major upgrades. It also instructs the agent to stop on regressions rather than quietly expanding the work into application changes.
For a team adopting the package, repository access and patch scope deserve separate decisions. Granting an agent visibility across an organization should not silently become permission to modify every repository it finds.
What counts as a completed update
The implementation calls for baseline and post-change tests inside the relevant built image, plus runtime comparisons and a smoke test. It requires reports to distinguish blocked work and weak test coverage rather than treating missing evidence as success.
Those requirements are useful, but they remain instructions that need checking in the actual output. A passing build does not establish that every important application path was exercised, and a generated rollout plan is still a plan.
An opened pull request is not a deployed update. The handoff should make it clear who reviews the evidence, who merges the change and who verifies the running artifact afterward. That is where this kind of agent workflow can become valuable: less time assembling the change, with a clearer record for the people responsible for shipping it.
I'm Paras. I work at the intersection of AI, engineering and distribution.
If you're building an AI product, I help explain technical workflows and launches in content developers and founders can understand and share.