Cognition's Devin Security 2.0 Hunts Bugs and Ships Tested Fixes Automatically

Cognition is expanding Devin into an enterprise security system that hunts vulnerabilities, validates exploitability, writes fixes, runs tests, and opens pull requests. Devin Security 2.0 combines high-capability security models, a threat model built with the customer’s team, and parallel cloud virtual machines that reproduce and patch bugs.

Through its launch offer, Cognition says it will allocate hundreds of millions of dollars in Devin credits. Enterprise customers can receive free deep scans, while new customers can receive remediation credits.

Security backlogs meet faster exploits

Recent disclosures from the PHP and Node.js projects show why Cognition is targeting vulnerability backlogs. Both projects reported remote code execution flaws, which can let an attacker run code on a vulnerable server. Signs of an underlying flaw in the PHP2Shell exploit chain had surfaced in 2024 without a concrete path to remote code execution. Wiz later reported that high-capability models helped its researchers develop and adapt working exploits across several applications within hours. Bugs that once appeared difficult to exploit may now become practical attack paths more quickly.

AI-assisted reporting is also increasing the defender’s workload. The curl project received 181 security reports during 2025 and 356 during the first ten months of 2026. Maintainer Daniel Stenberg has described the challenge as a combination of high volume and credible submissions. Cognition is positioning Devin between that expanding intake queue and the teams responsible for validation and remediation.

Inside the scan pipeline

Devin Security 2.0 begins with a shared threat model and ends with a validated finding ready for remediation. Its four stages are:

1. Build the threat model. Devin maps systems, services, hosts, network policies, security groups, and communication paths. The customer supplies context that source code cannot establish, including deployment regions, internet exposure, and internal-only services.
2. Map and inspect the repository. Cognition’s Agentic MapReduce architecture identifies routes, authorization boundaries, data-access patterns, and other security-sensitive code. It divides selected code into batches for parallel analysis by multiple models.
3. Aggregate and validate. The system deduplicates reports, connects findings that could form an exploit chain, and re-scores severity against the threat model. Higher-priority findings are tested in a sandbox running the application.
4. Send confirmed issues to remediation. Approved findings can move into parallel sessions that reproduce the vulnerability, prepare a fix, run tests, and open a pull request with supporting evidence.

Cognition describes the orchestration layer as model-neutral, allowing it to run Anthropic’s Claude Mythos alongside other high-capability cybersecurity models. Customers access those models through Devin without negotiating separate contracts with each provider.

A fix comes with proof

Each high-priority finding can launch an independent cloud virtual machine. Cognition says Devin can run thousands of these sessions in parallel, with each one reproducing the issue, writing a fix, rerunning the reproduction test and existing test suite, and opening a pull request that records the results.

Some confirmed issues can receive operational mitigations before a permanent code change is approved. With sufficient access, Devin can add a web application firewall rule, change a cloud security group, or remove public access from a service. For recurring bug classes, it can propose broader refactors such as centralizing authorization checks or parameterizing database queries across a codebase.

Every proposed severity change requires human approval. Devin preserves the original Common Vulnerability Scoring System score beside its adjusted score, and it never relabels vulnerabilities in the U.S. Cybersecurity and Infrastructure Security Agency’s Known Exploited Vulnerabilities catalog. Compliance evidence can be exported as Plan of Action and Milestones records and into governance, risk, and compliance tools, helping teams track FedRAMP and PCI DSS remediation deadlines.

Old findings get another pass

Devin can also ingest findings from existing scanners, bug-bounty programs, advisories, and exported reports. It checks them against the current codebase, removes duplicates, tests selected issues in a cloud virtual machine, and proposes fixes. That workflow can convert Snyk and Wiz findings, along with years of CSV exports, into a current remediation queue.

Anthropic’s Glasswing review illustrates the value of re-triage. Of 1,752 findings initially rated high or critical, 90.6% were valid vulnerabilities. The review confirmed the original high or critical rating for 62.4%. Accurate detection and accurate severity ranking require separate validation, especially as submission volume grows.

Review becomes the constraint

Security teams using this workflow would receive findings with reachability evidence, a working reproduction, test results, and a proposed change. Human reviewers retain authority over severity adjustments and code changes, so review capacity, repository ownership, and merge controls determine how quickly the generated work reaches production.

Free model-based scans also change which repositories teams can afford to audit. Organizations can apply the same process to lower-priority services and old backlogs, then measure whether the resulting pull requests reduce remediation time enough to justify the service after promotional credits expire.

Before granting access

Running Devin across repositories can require access to source code, continuous integration systems, test environments, and, for operational mitigations, cloud and network controls. A rollout should define:

• Permissions: Use separate, least-privilege credentials for scanning, pull requests, testing, and infrastructure changes.
• Data governance: Confirm how source code, logs, prompts, secrets, and model outputs are transmitted, retained, and shared with model providers.
• Quality gates: Set requirements for reproductions, regression tests, code-owner approval, and rollback plans.
• Operational limits: Establish concurrency caps, repository scope, escalation paths, and ownership for failed or ambiguous sessions.
• Economics: Track post-credit pricing, reviewer time, pull-request acceptance, regression rates, and cost per confirmed fix.

Cognition’s claim can be evaluated through concrete measures: confirmation rate, time to reproduce, time to merge, reviewer hours per fix, and production regressions. Those results will show whether parallel agents can turn growing security queues into changes that engineering teams can safely ship.