While the Pentagon was making headlines on May 1 by signing AI deals with eight vendors and pointedly excluding Anthropic, Anthropic itself was doing something arguably more consequential for the enterprise software industry. It opened the public beta of Claude Security to every Claude Enterprise customer.
Claude Security is built on Opus 4.7 and benefits from Mythos, the cybersecurity-specialized model whose existence is the reason the White House reopened talks with Anthropic last month. The product scans codebases, finds vulnerabilities that pattern-matching tools miss, generates patches, exports to Jira and Slack, and runs on a schedule. There is no API integration to build, no agent harness to wire up, no professional services engagement to scope. Enterprise customers go to claude.ai/security and start scanning.
For CISOs, AppSec leads, and procurement teams already paying for some combination of Snyk, GitHub Advanced Security, Veracode, Checkmarx, Endor Labs, and SonarQube, this is the most significant repricing event in the application security category since the SAST market formed. Here is what shipped, what it actually does, where it falls short, and how to think about it before your next AppSec renewal lands.
What Anthropic Actually Released
Claude Security exited closed preview on April 30, 2026, and is now in public beta for all Claude Enterprise customers globally. Coming-soon access is promised for Claude Team and Max plans. Pro and individual users do not get it.
The product has three layers worth understanding separately.
The reasoning layer: Opus 4.7 sits underneath, with Mythos providing specialized cybersecurity capability. Mythos is the model that, according to public reporting, has surfaced thousands of zero-day vulnerabilities in production codebases — the same capability that prompted the President to call it "a separate national security moment" and reopen talks with Anthropic CEO Dario Amodei after the Pentagon's blacklisting. Most of that capability now ships as a commercial product.
The product layer: A web UI at claude.ai/security and a sidebar entry point inside the existing Claude.ai chat interface. Connected GitHub repositories. Scheduled scans on a configurable cadence. Targeted directory scans within a repo. A multi-stage validation pipeline that reduces false positives by stacking confidence ratings before a finding surfaces to a human reviewer. Findings can be dismissed with documented reasons that persist across scans. Export formats include CSV and Markdown for audit systems. Webhook integrations to Slack, Jira, and other ticketing systems.
The integration layer: This is where it gets interesting. Anthropic announced that CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, Trend Micro, and Wiz are integrating Opus 4.7 capabilities into their existing security platforms. Accenture, BCG, Deloitte, Infosys, and PwC are building Claude-integrated solutions for vulnerability management, secure code review, and incident response. So Claude Security is not just a standalone product. It is also the engine that the major SOC and CSPM vendors are quietly putting under their existing AppSec and detection products.
The pricing posture is the most disruptive part. For Claude Enterprise customers, Claude Security is included in the existing subscription. The marginal cost of running scans is metered against your existing Claude usage at the underlying token rate, which works out to roughly $15-$25 per single-codebase review depending on size and depth. For an enterprise already paying for Claude as their assistant of record, security scanning is now closer to a feature toggle than a procurement event.
What It Does That Existing Tools Cannot
The honest version of the AppSec market in 2025 was that you needed three tools to cover the surface area, and even then you missed half the real vulnerabilities.
Snyk was strong on software composition analysis — finding known CVEs in your open-source dependencies and container images. GitHub Advanced Security gave you secret scanning, dependency review, and rule-based SAST inside the GitHub workflow. Veracode and Checkmarx covered enterprise-grade SAST and DAST with compliance-acceptable reporting. Sonarqube did code quality and security rules. Endor Labs differentiated on reachability analysis to cut SCA noise. None of these tools — by design — could find a complex business logic flaw, a cross-component data flow vulnerability, an authentication bypass that emerged from how three services interacted, or a privilege escalation that required reasoning about state across the codebase.
The bet underneath Claude Security is that those second-class problems — the ones a senior application security engineer finds during a manual code review — are exactly what a frontier reasoning model can find. The product is positioned not as "another SAST tool" but as "the layer that catches what SAST cannot."
In practice that means two specific capabilities.
First, whole-codebase reasoning. Claude Security does not scan files in isolation. It traces data flows across files, reads how components call each other, and synthesizes the network effect of changes. The classic example is a SQL injection that exists not in one query but in the combination of an upstream input validator that allows certain characters and a downstream query builder that does not parameterize a specific column. Pattern matching cannot see that. Reasoning over the whole codebase can.
Second, patch generation, not just detection. Every legacy AppSec tool tells you there is a vulnerability and gives you a Jira ticket. Claude Security tells you there is a vulnerability and proposes the diff to fix it. For mid-sized engineering teams, that closes the most expensive part of the AppSec lifecycle: the gap between "we found a critical" and "we shipped the fix."
The track record claim from Anthropic is that "hundreds of organizations have used it to discover and fix exploits in production code, including vulnerabilities existing tools had missed for years." That is a strong claim. It will be tested in public over the next two quarters as enterprise customers run Claude Security alongside their existing stacks and report what each tool found and missed.
Where It Falls Short
Three specific limitations are worth knowing before you build it into your AppSec strategy.
GitHub-only at launch. Claude Security currently supports only GitHub-hosted repositories. If you are on GitLab, Bitbucket, Azure DevOps, AWS CodeCommit, or self-hosted Git, you wait. For regulated enterprises that mandate self-hosted source control for compliance reasons, this is a hard block until VCS support broadens.
SCA is not the focus. Claude Security is not a replacement for software composition analysis. If you turn it on and turn off Snyk, you stop seeing known CVEs in your open-source dependencies. The right framing is that Claude Security catches reasoning-class vulnerabilities; you still need a CVE database and dependency scanner for the known-knowns.
Beta means beta. Public beta in 2026 means: feature gaps will be filled by the GA, support is "best effort," SLAs do not apply yet, and enterprise change management cannot treat it as a system of record for compliance audits today. Use it to find vulnerabilities and ticket fixes; do not yet use it as the primary evidence in your SOC 2 or ISO 27001 audit.
There is also a less-discussed concern that matters for regulated industries: data egress. Claude Security needs to read your source code to reason over it. Anthropic has not, as of the public beta launch, published the specific data handling, retention, residency, and customer-data-isolation guarantees at the level of detail that defense, financial services, and healthcare CISOs will require. If your code is regulated, get those terms in writing before you connect production repositories. Beta launches are exactly when those guardrails get negotiated.
The Real Disruption Is Bundling
The strategic question Claude Security poses is not "is it better than Snyk." It is "what happens to the AppSec budget line when the assistant your developers already use also scans the code they write?"
Three structural shifts to model.
Shift one: the bundled-in case kills standalone budget. When AppSec scanning shows up at no marginal procurement cost inside the Claude Enterprise subscription you already signed, the standalone Snyk renewal becomes a harder conversation. The CFO question is no longer "is Snyk worth $X per developer per year?" It is "is Snyk worth $X per developer per year on top of what we already get from Claude?" That math changes the renewal.
Shift two: the major security vendors are choosing to be hosts, not competitors. CrowdStrike, Palo Alto Networks, SentinelOne, Trend Micro, and Wiz are integrating Opus 4.7 into their own platforms. They are betting that their distribution, their existing buyer relationships, their compliance attestations, and their ability to deliver an end-to-end SOC story is more durable than building competing reasoning capability. That is the same posture banks took with cloud — host the disruptor inside your perimeter rather than fight it. It will work for the platforms with strong distribution. The standalone AppSec vendors who do not get bundled in are exposed.
Shift three: the consultancy partnership tier is the leading indicator. Accenture, BCG, Deloitte, Infosys, and PwC each signed on as deployment partners. That tells you where the dollar volume will land first: large regulated enterprises with multi-year transformation programs, where the consultancy-led delivery model is how AI security gets actually deployed at scale. If you are at one of those firms and you have an AppSec transformation on the books, expect the integrator pitch to lead with Claude inside the next 90 days.
The losers in this picture are the standalone SAST tools that staked their differentiation on rule-based scanning of complex codebases. Their moat — proprietary rule libraries built up over a decade — is exactly what a frontier reasoning model commoditizes. The winners are the platforms that own distribution, the SCA specialists who do something fundamentally different (dependency CVEs and reachability analysis), and the integrator-led delivery model that turns Claude Security into a deployed program inside an enterprise.
The bigger context: this is Anthropic's third major enterprise-stack move in six weeks. Glasswing in mid-April established the cybersecurity partnership tier with the major SOC vendors. The Mythos disclosure brought the cyber capability into the political conversation and forced the Pentagon to negotiate. Claude Security on May 1 turns the underlying capability into a sellable enterprise product on standard commercial terms. Each move pushed Anthropic deeper into the enterprise security buyer's wallet — first as a partner inside other vendors' products, now as a directly-billable feature inside an enterprise subscription. The pattern is unmistakable: Anthropic is building an enterprise security business in the open, one quarter at a time, and the AppSec category is the first one being repriced.
What CISOs and AppSec Leaders Should Do This Week
Three concrete actions.
One: turn on the beta on a non-production repository this week. If you have a Claude Enterprise contract, the cost of trying Claude Security is essentially zero. Pick a recently-built service, run a scheduled scan, and compare the findings against what your incumbent SAST tool produced over the last quarter. The empirical answer to "is this real" is two scans away. Do this before your next renewal cycle so the data is fresh when the procurement conversation starts.
Two: pull your AppSec contract terms now and check the auto-renewal dates. If your Snyk, Veracode, GitHub Advanced Security, or Checkmarx contract renews in the next two quarters, you want negotiation leverage. The moment Claude Security demonstrates parity-or-better on reasoning-class vulnerabilities in your environment, your incumbent's pricing becomes negotiable. Get into that conversation with data, not opinion.
Three: get the data-handling terms in writing before connecting production code. For regulated industries, treat Claude Security like any new data processor. Get the SOC 2 / ISO 27001 attestations, the data residency guarantees, the customer-data-isolation specifics, and the deletion SLAs in writing under your master agreement. Beta is exactly when vendors are most flexible on these terms. Use that.
The Pentagon excluded Anthropic on usage-policy grounds and now wants Mythos enough to negotiate around the broader dispute. The commercial market just got the same capability shipped as a beta inside the Claude Enterprise subscription that hundreds of large organizations already pay for. The DoD and the Fortune 500 are about to be running the same underlying model against very different threat surfaces, on very different procurement terms. Watch the AppSec budget line over the next two quarters. That is where the repricing will be most visible.
If your organization is rebuilding its AppSec program for the AI era — or trying to figure out which incumbents to keep, which to consolidate, and which to displace — the architectural choices made in the next two quarters will set the cost basis for the rest of the decade. Run the beta. Pull the contracts. Make the call with data.
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.
Continue Reading
- EPAM Just Bet 10,000 Architects on Claude. Here’s Why.
- [Anthropic Kills Seat Fees: Pay $0.004/1K Tokens vs OpenAI $20/Month](/article/anthropic-usage-based-pricing-2026)
- OpenAI Guarantees 17.5% Returns to PE Firms in $10B AI Deal
