Devin Security Threats Turns New CVEs Into Pull Requests
Two nasty vulnerabilities dropped in the same week. Java4Shell in the Java runtime, NGINX2Root in nginx. Both can end in remote code execution. And Cognition's answer isn't a blog post telling you to patch — it's a product that finds the vulnerable code in your repos and writes the patch for you.
That's Devin Security Threats, launched November 19, 2026. It's a managed catalog of newly disclosed vulnerabilities wired straight into the Devin coding agent. You open a threat, you scan your connected repos, Devin triages what it finds and prepares fixes as pull requests.
The honest framing up front: I haven't used it. It's brand new, and the launch materials leave real questions open — price, validation guarantees, how good the generated patches actually are. Everything below is the launch as reported by AlphaSignal, plus what you can do with it today, including the part Cognition open-sourced. No hands-on testing was possible for this article.
What it actually is
A threat record describes the affected software, the attack path, the expected impact and the available remediation. Cognition publishes records for major CVEs. Your own security team can also create entries for private findings — bug-bounty reports, incident reviews, patterns specific to your codebase.
Each threat card has a "Scan with Devin" button. Hit it and Devin sweeps your connected repositories using the same Agentic MapReduce process as the rest of Devin Security: the investigation splits across parallel agent sessions, searches the relevant files and repos, and consolidates everything into one set of findings.
The workflow around it is the part worth understanding, because it's more than a scan button:
- Manual scanning. Launch a scan from any individual threat record.
- Automatic triage. Pick your repos and a minimum severity, and Devin scans them whenever a qualifying threat is published. No one has to notice the CVE first.
- Prioritised findings. Results are ranked against your team's threat model, and you can assign findings to Devin for remediation.
- Pull-request delivery. Devin prepares code changes as pull requests for review, testing and approval.
- Automatic dismissal. Threats with no findings in your repos close themselves, so the queue doesn't fill with noise.
- Custom detection. Internal threat records can target proprietary code patterns, not just public CVEs.
The two launch threats are the ones that make the timing real. Java4Shell (CVE-2026-114938) is a Java runtime flaw that can expose memory or enable remote code execution. NGINX2Root is five memory-corruption flaws in nginx that can also lead to remote code execution. Java and nginx sit underneath a huge share of production systems, so either disclosure on its own would trigger a round of repo reviews. Both landing in the same week is exactly the operational problem Cognition is selling against: finding the affected code fast enough to start fixing it before exploits circulate.
I covered Java4Shell's patch separately on November 18, and Devin Bots — the AI-manager layer this builds on — on November 17. This is a different story: not the vulnerability, and not the manager, but the pipeline from disclosure to pull request.
The part you can use today, for free
Here's the bit that matters most if you don't use Devin at all. Cognition open-sourced two skills in its Vulnerability Fixes plugin:
-
fix-nginx-vulncovers the five NGINX2Root CVEs -
fix-java-vulnhandles Java4Shell (CVE-2026-114938)
Install them with:
devin plugins install CognitionAI/devin-marketplace#plugins/devin-vuln-fixes
The skills are open-source Markdown files. That means you can read the remediation steps before trusting them, adapt them as internal checklists, or hand them to a different coding agent entirely — Cognition says they work with Claude Code and Cursor too.
That's the practical payload of this launch, and it's the reason I'm writing about it rather than just relaying the announcement. A security announcement you can only watch is a press release. A fix procedure you can read, run in your own agent, and argue with is a tool.
If you run Claude Code or Cursor and either of these CVEs touches your stack, the honest move is to pull the skill file, read it, and run it against a copy of the affected repo first. Don't let any agent — Devin's or anyone else's — touch your main branch on a security fix you haven't read.
Where the fixes stop
Devin prepares the pull request. That's where it stops. Code review, test execution, deployment approval and production validation stay with your team. Cognition is explicit about that, and it's the right line — a generated patch for a memory-corruption flaw is exactly the kind of change that deserves a careful human diff review, not a fast merge.
The boundaries go further than that. Devin can only inspect the repositories and context it has access to. Your team still owns asset inventory, tracking which versions are actually deployed, runtime mitigations, dependency upgrades, rollout planning and post-deployment verification. A scan that finds nothing in the repos Devin can see is not proof that nothing is exposed in production.
And there are things Cognition hasn't disclosed at all: no pricing, no list of supported repository hosts, no scan-duration targets, no validation guarantees for generated patches. If you're evaluating this, those are the questions to ask before anything else, along with the ones the launch materials themselves flag — repository permissions, how generated code gets reviewed, secret handling, CI integration, and the false-positive and false-negative rates on your own codebase.
The catch: faster discovery means faster response is now your problem
Cognition's argument for why this exists now is worth taking seriously, even if you're sceptical of the product. AI-assisted security research is increasing the rate at which vulnerabilities surface. Fuzzers probe software with malformed inputs; agents can inspect and test many candidate paths across mature codebases. Java4Shell and NGINX2Root in the same week are the current examples, not an exception.
When a new nginx CVE lands, the work is: locate the affected code, work out which versions are deployed, prepare patches, test dependent services and coordinate rollout — all before reliable exploits circulate. Discovery is getting faster. Response capacity is the constraint, and it's the part that doesn't scale by itself.
That's the real shift here. The bottleneck in vulnerability response used to be finding the bug. Now the bugs are being found faster than most teams can patch them. A product that turns a disclosure into a pull request is trying to close that gap — but it only closes the middle of the pipeline. The ends, inventory and verification, are still yours.
Should you care yet? A quick checklist
If Java or nginx runs anywhere in your stack: check the two threat entries now. Both are live in Devin's security dashboard. Even if you never touch Devin, read the open-source fix skills — they're the fastest way to see what the remediation actually involves.
If you already use Devin: this is your announcement. Set up automatic triage with a minimum severity and let qualifying threats scan your repos as they're published, rather than waiting for someone on the team to notice the CVE.
If you use Claude Code or Cursor: the skills are portable. Pull the Markdown, read it, and use it as a checklist for your own agent. You don't need a Devin account to benefit from the part Cognition open-sourced.
If you lead engineering: the question isn't whether to use this specific product. It's whether your team could answer, today, which repos run the affected Java or nginx versions and how long a patch would take to reach production. If that answer takes a day of grep and Slack threads, the gap this product targets is your gap too.
What we still don't know
- The price. Nothing published.
- Supported repository hosts. Not specified in the launch materials.
- Scan speed. No duration targets published.
- Patch quality. No validation guarantees, no false-positive or false-negative rates.
- How the generated patches hold up in review. The first teams to merge Devin-prepared security fixes will tell us more than any launch post can.
Our take
The product is a promise, not a result — same as every launch-day story. But the open-source skills are a result, and they're available right now. That split is what makes this launch worth more than a skim: the managed catalog is a bet on Cognition's roadmap, while the fix skills are something you can inspect and use today, in whatever agent you already run.
The test that matters is simple and it hasn't happened yet: a team takes a Devin-prepared patch for one of these CVEs, reviews it, tests it, ships it, and tells us whether the diff was clean. Until then, treat the pull requests like any other agent-written code — read every line, run the tests, and don't merge on trust.
If either of these CVEs touches your stack, have you checked which repos are actually affected yet — and what would your patch timeline look like if a third one landed tomorrow?
Sources: AlphaSignal, "Cognition's Devin Turns Security Vulnerabilities Into Auto-Generated Code Fixes," November 19, 2026 (https://alphasignal.ai/news/cognition-s-devin-turns-security-vulnerabilities-into-auto-generated-code-fixes). Technical details for the underlying vulnerabilities are in the Java4Shell disclosure (https://calif.io) and the NGINX2Root disclosure (https://xint.io). Both threat entries are live in Devin's security dashboard (https://app.devin.ai). Cognition has not disclosed pricing, supported repository hosts, scan-duration targets, or validation guarantees for generated patches — where this article says a fact is unknown, that is because it has not been published, not because it was missed. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links. AI disclosure: this article was drafted with AI assistance and checked against the AlphaSignal report.
Also on ToolsDig: Devin Security Threats Turns New CVEs Into Pull Requests
Top comments (0)