On October 26, 2026, Cognition, the company behind the Devin coding agent, announced a major upgrade to Devin Security and a pledge of "hundreds of millions of dollars in Devin credits" to help enterprises find and fix their most critical vulnerabilities. The headline features are deep scans powered by Claude Mythos and other frontier cyber models, and free scans for enterprise customers. The details are in Cognition's announcement.
This post sticks to what the announcement actually says, adds the context from Cognition's earlier Security Swarm launch, and lists the questions to ask before accepting free credits from any vendor. It is not a hands-on test: we have not run Devin Security, and the performance numbers cited below are Cognition's own.
TL;DR
| Question | Answer |
|---|---|
| What launched? | A major upgrade to Devin Security: deep scans, a living threat model, exploitability-based rescoring, and multiple remediation paths |
| Which models? | Claude Mythos and "other frontier cyber models"; Cognition calls the product model-neutral |
| Who gets free scans? | Current and future enterprise customers, per Cognition |
| Who gets remediation credits? | New customers can also receive them |
| How much? | "Hundreds of millions of dollars" in Devin credits in total; no per-customer figure published |
| How do I start? | Contact Cognition; there is no self-serve sign-up in the announcement |
| What is independently verified? | Nothing yet; the announcement is a vendor post |
What Cognition actually announced
The post describes "two initiatives." The first is the upgrade itself. Devin Security is described as "our product for finding, prioritizing, and remediating vulnerabilities end to end," and now offers deep scans powered by Claude Mythos and other frontier cyber models. Cognition stresses that enterprise security teams can use these capabilities "through Devin without needing separate access to the models." That matters because access to the strongest cyber models has been tightly controlled: Anthropic first limited its Mythos class to select organizations, a story we covered in Claude Mythos Preview and Project Glasswing and again when Mythos 5 security scans arrived with a defender advantage fund.
The second initiative is money. Cognition says it is "committing hundreds of millions of dollars in Devin credits" and offering "current and future enterprise customers free scans with frontier cyber models to find complex vulnerabilities." New customers "can also receive remediation credits to help get their most critical flaws fixed before attackers can exploit them." The company adds that its team "will work alongside customers through investigation and remediation."
Credits are not cash, and a figure measured in credits depends on how Cognition prices them internally. A headline of hundreds of millions in credits therefore says more about the scale of the go-to-market push than about what a given security team will receive. The post publishes no cap, expiry or eligibility threshold, so treat the number as an intent signal, not a spec.
How Devin Security says it works
The announcement describes a pipeline in four stages.
1. A living threat model
Devin builds a "living threat model" with your security and engineering teams. It maps systems, services, hosts, how they communicate, network policies and security groups, how code is deployed and which code is unused. Humans fill the gaps Devin cannot infer from code, such as which region a service runs in or whether it is internal-only. Devin also asks about business impact: what happens if one user can read another's data, gain administrator access or use the product without paying. Those answers feed later scans and prioritization.
This is the part that differs most from a classic scanner. A tool that assigns severity without knowing the deployment can overstate some risks and understate others, and Cognition argues those ratings still "trigger compliance or policy deadlines." A threat model is the mechanism it proposes to fix that.
2. Deep scans
Deep scans are "a new way to uncover complex chained exploits across large codebases." The security harness uses the threat model to guide the scans, investigating how a flaw could be exploited and not only whether a pattern matches. Devin also ingests findings through APIs and integrations from existing scanners, bug bounties and advisories, removes duplicates and re-verifies them with frontier models.
3. Rescoring by exploitability
Devin re-scores findings, its own and other tools', against the threat model according to exploitability and potential damage in your environment. Teams see which flaws need fixing now, which can be scheduled, and which warrant a documented risk adjustment.
4. Four ways to resolve a finding
Cognition lists four responses, since "resolving a finding does not always require a pull request":
| Response | What Devin does | Guardrail |
|---|---|---|
| Relabel with evidence | Documents false positives or proposes risk adjustments based on reachability and access controls; can prepare risk-acceptance records | Each proposal needs human approval; the original CVSS score stays visible; KEV-listed vulnerabilities are never relabeled |
| Mitigate in infrastructure | With configured access, applies a WAF rule, changes a security group or removes public access | Reduces exposure while a permanent fix is prepared; the flaw stays open |
| Fix the flaw | Runs thousands of parallel sessions in independent virtual machines, each reproducing the bug, writing a fix, running the reproduction and existing tests, and opening a PR with evidence | PRs go to review |
| Refactor out the class | Estimates and helps implement a broader change, such as centralizing authorization checks or parameterizing queries | Larger change, human-led |
Results sync back to existing security tools, and supporting records can be exported to POA&M and GRC tools, which are the plan-of-action and compliance workflows many regulated teams already run.
Where Security Swarm fits
Devin Security Swarm launched on July 1, 2026. Cognition described a swarm of parallel agents that each reason across a segment of the codebase, catching business-logic flaws, chained auth bypasses and cross-service exploit paths, then composing findings into attack paths and reproducing each in an isolated sandbox. Only confirmed-exploitable vulnerabilities reach the security team, and Devin writes the patch.
Cognition published a benchmark with that launch: 50 real-world vulnerabilities, each tied to a published GitHub Security Advisory, across repositories in more than a dozen languages.
| Harness | Recall | Cost per run |
|---|---|---|
| Devin Security | 72% | $90.23 |
| Claude Security | 68% | $131.87 |
| Codex Security | 48% | $118.20 |
| Cursor Security | 26% | $4.60 |
Source: Cognition's Security Swarm launch post, July 1, 2026. Self-reported, not independently verified.
Read this table with care. It was produced by the vendor, on a 50-item set the vendor selected, and the cheapest tool in it is also the least accurate. A 4-point recall gap between the top two is small on 50 items. It tells you Cognition's harness is competitive in Cognition's test, not that it wins on your code. The October 26 post, notably, includes no new benchmark for the Mythos-powered deep scans, so the upgrade's headline capability has no published numbers yet.
Why Cognition is making this move now
A gear with a green bandage over a crack, standing for automated patching in Devin Security
The announcement is blunt about the threat model: "AI is making sophisticated attacks easier to carry out and giving defenders less time to respond." That is consistent with a run of stories this month. The PHP2Shell and Patch the Planet effort showed AI-found bugs being patched at scale, Trail of Bits examined regressions that fast patching can introduce, and Bank of America put 1,000 contractors on Mythos-found flaws. The bottleneck in all three is the same: finding is getting cheap, fixing is not.
Cognition's pitch targets that bottleneck. Its post notes that security teams "face a growing volume of findings and finite engineering capacity," and that Devin's cloud-native design, with one virtual machine per session, lets it run thousands of scans and fixes at once. For Cognition the offer also fits a growth stage: the company's own blog lists a Series E of over $2B at a $48B valuation on September 8 and a $1B annualized revenue run rate on September 25, plus the SWE-2 coding model and Fusion harness in September. Security is a natural enterprise wedge for a company that sells agent compute: every remediation session consumes credits, and free scans create the findings that need fixing.
That is not a criticism, only a reason to read the offer the way a procurement team would.
How it compares with other AI security tooling
The market now has several routes to the same capability.
- Anthropic's own product. Claude Security went to public beta in July (our coverage) and later added continuous scanning, fix tracking, an API and MCP support (details). Anthropic also offers free AI vulnerability scans to open-source maintainers. If your goal is Mythos-class scanning of a repository, going to the model lab directly is an option, as is a reseller layer such as Devin.
- OpenAI-side tooling. Cloudflare's Daybreak work with Codex shows an alternative stack for discovery plus remediation.
- Devin's differentiator. On Cognition's own account, it is the combination of a threat model, cross-tool ingestion, and parallel VM-based fixing, delivered as a managed, model-neutral product with services attached.
The trade-off is typical for a managed layer. You get convenience, one contract and an integrated fix workflow; you give up some visibility into which model found what, and you add another vendor that touches your source code.
What to ask before you accept free credits
A sealed envelope with a small green padlock, representing careful disclosure terms for AI vulnerability scanning
Free scanning on frontier cyber models is attractive, and the questions below apply to any similar offer.
- What exactly is free? Scans of how many repositories, lines or hours? What happens when the free scope ends, and does remediation convert to paid use automatically?
- Where does my code go? Which models see it, under which provider's retention and training terms, and in which region? "Model-neutral" means the answer can change as Cognition swaps models.
- What is the access scope? Infrastructure mitigations such as WAF changes need configured access. Start read-only, and require approval for anything that changes production.
- How are findings validated? Ask for the reproduction evidence Cognition says accompanies each PR, and spot-check a sample against your existing scanner and a human review.
- Who owns disclosure? If a scan finds a flaw in a third-party dependency, who reports it and how?
- What is the false-positive rate on your code? Run a bake-off on a known-vulnerable service, with your own seeded bugs, before widening scope.
- How are rating changes audited? Cognition says the original CVSS stays visible and KEV items are never relabeled; confirm your GRC tooling records who approved each adjustment.
- What is the exit plan? Confirm you can export findings, threat models and evidence in a portable format.
What this means for builders and security teams
If you ship software at an organization with a security team, the practical changes are small but real.
- Backlog math is shifting. If scans are free, the cost center moves to review and merge capacity. Plan for more fix PRs, not just more findings.
- Context beats severity. The threat-model approach is worth copying even without Devin: documenting which services are internal-only, which code is dead and what a breach would cost turns a flat CVSS list into a ranked queue.
- AI-authored fixes need the same review as AI-authored code. Regression risk is real, and the Trail of Bits piece linked above is a good reminder. Require tests with every automated fix.
- Do not hand agents more permissions than the task needs. The same agents that patch can misfire. If you let coding agents touch infrastructure, put an approval layer in front of irreversible actions; AgentBeam, from the explainx.ai team, is one option for stopping agents before they take dangerous actions.
What is confirmed and what is not
| Claim | Status |
|---|---|
| Devin Security upgraded, with deep scans on Claude Mythos and other frontier cyber models | Stated by Cognition in its October 26 post |
| Hundreds of millions of dollars in credits | Cognition's figure; no total, cap or per-customer amount published |
| Free scans for current and future enterprise customers; remediation credits for new customers | Stated by Cognition; terms not published |
| Security Swarm: 72% recall at $90.23 per run on 50 advisories | Cognition's own benchmark from July 1; not independently verified |
| Deep-scan accuracy or cost with Mythos | No published numbers |
| Independent reviews or customer results for the October upgrade | None found at the time of writing |
How to check it yourself: read the announcement and the Swarm post, ask the sales team for the credit terms in writing, and run a trial against a service whose vulnerabilities you already know. Fund a result you can measure, not a headline.
Related reading
- Claude Mythos Preview and Project Glasswing
- Claude Mythos 5 security scans and the defender advantage fund
- Claude Security in public beta
- Claude Security continuous scanning, fix tracking, API and MCP
- Anthropic's free open-source vulnerability scanner
- PHP2Shell and Patch the Planet
- Trail of Bits on patching bugs and regressions
- Cognition SWE-2 and Devin
Information is current as of October 26, 2026. Product terms, credit eligibility and model availability can change; confirm with Cognition.