Cognition Puts Anthropic's Restricted Mythos 5.2 Inside Devin Security

Cognition’s announcement adds Claude Mythos 5.2 to Devin Security for eligible enterprise customers. The company also committed more than $100 million in scan credits to help organizations investigate and remediate existing vulnerability backlogs.

Devin Security keeps the restricted model behind a managed interface that scans customer-owned code, validates potential exploits and returns findings or remediation pull requests. Customers gain access to Mythos-class security analysis without receiving the model itself or obtaining a separate direct entitlement from Anthropic.

Anthropic’s restricted cyber model

Anthropic positions Mythos above its Opus tier for cybersecurity and biology research, according to Anthropic’s model page. Access runs through a vetted partner program because the model can perform advanced vulnerability research and exploit development.

Anthropic says Mythos became the first known AI model to compromise a simulated corporate network autonomously from initial access through completion. The company also credits it with finding more than 2,000 previously unknown vulnerabilities in production software within two months.

Anthropic limits distribution through Project Glasswing and managed products such as Claude Security. These services expose specific defensive workflows and return structured results while withholding direct API and model access. Devin Security now follows that pattern.

How Cognition reached 73.5

Cognition evaluated the models against 25 repositories containing 1,781 publicly reported vulnerabilities. The test set covered C, Python, Go, PHP and TypeScript projects using Django, Flask, Laravel and Next.js. Each repository was pinned to a revision that still contained the known flaws.

The F2 metric combines precision, the share of reported findings that are correct, with recall, the share of known vulnerabilities found. Its formula gives recall four times the weight of precision. Cognition ran each model three times on every repository and reported the mean, reducing the effect of variation between runs.

Cognition’s earlier Swarm evaluation asked whether a model found one target vulnerability in each repository and provided little information about false positives. The new test covers every known vulnerability in the selected revisions, giving teams a broader measure of detection quality. The results remain an internal vendor benchmark and have not been independently reproduced.

Credits aimed at security backlogs

The credit program subsidizes scans for enterprise customers with accumulated vulnerabilities across large codebases. Cognition has announced the total value of the pool, while eligibility, per-customer allocations, expiration dates and other commercial conditions depend on the application process.

Anthropic’s disclosures and a separate Bain analysis describe Mythos as capable of making advanced exploit development accessible to users with limited security experience. Reported examples include a 20-gadget return-oriented programming chain, which assembles an exploit from executable snippets already present in a binary. Wider access to such capabilities raises the value of finding and patching exposed flaws before attackers can automate the same work.

Devin Security Swarm applies the model through a managed remediation workflow:

1. Parallel agents divide the codebase into smaller areas for investigation.
2. The agents trace data flows and combine findings into potential attack paths.
3. Runtime checks determine whether suspected vulnerabilities are exploitable.
4. The system prepares remediation pull requests for engineering review.

Traditional static application security testing examines source code for risky patterns, while dynamic testing probes a running application. Devin Security combines code analysis, exploit-path reasoning, runtime validation and patch generation in one workflow, which can reduce the manual triage required after a scan.

The swarm stays in place

Cognition built Security Swarm around an architecture called Agentic MapReduce. Multiple agents inspect separate parts of a repository in parallel, then aggregate their evidence to identify vulnerabilities that span files, services or components.

In Cognition’s earlier 50-vulnerability benchmark, Security Swarm found 36 targets at a reported cost of $90.23 per run, outperforming Claude Security and Codex Security on recall and cost. That evaluation used a different task set and scoring method, so its results cannot be compared directly with the new F2 scores. The Mythos 5.2 integration preserves the parallel workflow while supplying a stronger model for vulnerability and exploit reasoning.

Questions to settle before deployment

• Access: Eligible enterprise customers can request access; Cognition has not published universal qualification criteria.
• Coverage: The benchmark establishes results for five languages and four web frameworks. Teams need a complete support matrix for other stacks, monorepos and generated code.
• Precision: The recall-heavy F2 score does not disclose per-language false-positive rates or performance on repositories outside Cognition’s test set.
• Data handling: Buyers need written terms covering source-code retention, deployment region, logging, model training and the treatment of embedded secrets.
• Patch safety: Generated pull requests still require tests, regression checks and human approval before deployment.
• Credit terms: The practical value of the subsidy depends on scan pricing, allocation limits, expiration dates and whether rescans consume additional credits.

Managed distribution gives Anthropic and Cognition control over model access while exposing a focused defensive workflow to enterprise teams. Adoption will depend on independent validation, operational fit and contract terms, but the release shows how restricted security models can reach developers through scanning and remediation products.