Skip to main content
Back

Cognition's Devin Turns Security Vulnerabilities Into Auto-Generated Code Fixes

Cognition's new Devin Security Threats feature turns fresh CVE disclosures into one-click scans and auto-generated pull requests across your repos.

  • Cognition launched Devin Security Threats, a managed CVE index with one-click agent scanning and remediation.
  • Java4Shell (Java runtime RCE) and NGINX2Root (five nginx memory bugs) are the first two indexed threats.
  • Scans use Devin's Agentic MapReduce to fan out across repos, then rank findings against your threat model.
  • Auto-triage mode scans selected repos for every new threat and opens pull requests automatically.
  • New fix-nginx-vuln and fix-java-vuln skills ship in the Vulnerability Fixes plugin.
  • Skills are open-source Markdown playbooks usable with any coding agent, not just Devin.

Cognition has launched Devin Security Threats, a managed catalog of newly disclosed vulnerabilities connected to its Devin coding agent. Developers can scan connected repositories from each threat entry, triage findings, and ask Devin to prepare fixes as pull requests.

Two disclosures, one response queue

The release arrives alongside two serious disclosures: Java4Shell, a Java runtime vulnerability that can expose memory or enable remote code execution, and NGINX2Root, a group of five memory-corruption flaws in nginx that can also lead to remote code execution.

Threat Affected software Potential impact Identifier
Java4Shell Java runtime Memory disclosure and remote code execution CVE-2026-114938
NGINX2Root nginx Memory corruption and remote code execution Five CVEs

Java and nginx underpin many production systems, so either disclosure can trigger reviews across numerous services and repositories. Releasing both in the same week illustrates the operational problem Cognition wants to address: identifying affected code quickly enough to begin remediation before exploit activity expands.

Threat records become repository scans

In Devin’s terminology, a threat record describes the affected software, the attack path, the expected impact, and available remediation. Cognition publishes records for major CVEs, while internal security teams can create entries for private findings such as bug-bounty reports, incident reviews, and organization-specific code patterns.

Devin Security dashboard showing NGINX2Root and Java4Shell threat cards with Scan with Devin buttons
Each threat card can start a scan across repositories connected to Devin.

Selecting Scan with Devin starts a repository sweep using the same Agentic MapReduce process as the rest of Devin Security. The system divides the investigation among parallel agent sessions, searches relevant files and repositories, and consolidates the results into one set of findings.

A shorter path from finding to patch

Devin turns a published threat into a repeatable detection and remediation workflow:

  • Manual scanning: A developer can launch a scan from an individual threat record.
  • Automatic triage: Teams select repositories and a minimum severity, then Devin scans them when qualifying threats are published.
  • Prioritized findings: Results are ranked against the team’s threat model and can be assigned 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 can be closed automatically to reduce queue noise.
  • Custom detection: Internal threat records can target proprietary code patterns as well as public CVEs.

Generated fixes stop at pull requests, leaving code review, test execution, deployment approval, and production validation with the development team.

Interactive fixes stay portable

Cognition has also added two skills to its Vulnerability Fixes plugin. The fix-nginx-vuln skill covers the five NGINX2Root CVEs, while fix-java-vuln handles Java4Shell, identified as CVE-2026-114938.

Developers can install the plugin from the Devin Marketplace or through the command line:

devin plugins install CognitionAI/devin-marketplace#plugins/devin-vuln-fixes

The skills are open-source Markdown files, so teams can inspect the remediation steps, adapt them as internal checklists, or provide them to other coding agents such as Claude Code and Cursor.

Faster discovery raises the response burden

Cognition argues that AI-assisted security research is increasing the rate at which vulnerabilities surface, with Java4Shell and NGINX2Root serving as current examples. Fuzzers probe software with malformed or unexpected inputs, while agents can inspect and test many candidate paths across mature codebases.

As discovery accelerates, response capacity becomes a constraint. A newly published nginx CVE can require teams to locate affected code, determine which versions are deployed, prepare patches, test dependent services, and coordinate rollout before reliable exploits circulate.

Repository scans have clear boundaries

Devin can inspect only the repositories and context available to it. Teams remain responsible for asset inventory, deployed-version tracking, runtime mitigations, dependency upgrades, rollout planning, and post-deployment verification.

Teams evaluating the service will also need to assess repository permissions, generated-code review, secret handling, CI integration, and the false-positive and false-negative rates for their codebases. Cognition’s launch materials covered here do not specify pricing, supported repository hosts, scan-duration targets, or validation guarantees for generated patches.

Availability and source material

Cognition says both threat entries are live in the security dashboard. Technical details for the underlying vulnerabilities are available in the Java4Shell disclosure and the NGINX2Root disclosure.

The product pairs a managed vulnerability feed with an agent that can inspect a team’s code and prepare changes. Its practical value will depend on scan accuracy, repository coverage, patch quality, and how cleanly the generated pull requests fit into existing review and deployment controls.