Claude Security Is Live. Here's Your Move.
Anthropic shipped Claude Security in public beta on April 30. See the enterprise vulnerability scanner Claude Enterprise customers can deploy this week.
On April 30, Anthropic moved Claude Security from closed preview to public beta for Claude Enterprise customers. It runs on Claude Opus 4.7. It scans your codebase, finds real vulnerabilities, and writes patches for human review. Closed-preview customers reported Claude finding bugs that existing static analysis tools had missed in production code for years.
The thing that should land hardest for an enterprise reader: there is no API integration to build. No new contract to negotiate. If you already have a Claude Enterprise seat, you can request beta access today and start scanning before Friday.
I’ve been writing about restricted-access security AI for the last three weeks. GPT-5.4-Cyber gated to vetted defenders. Claude Mythos locked behind Project Glasswing partner agreements. The pattern was that frontier security capability was something you applied for and waited on. Claude Security is the first frontier security AI you can deploy this week with the contract you already have.
Quick Verdict
| The Move | What It Means for You |
|---|---|
| Claude Security public beta (Apr 30, 2026) | Available to all Claude Enterprise customers, no waitlist |
| Built on Claude Opus 4.7 | Same model that ships Anthropic’s design tool, now reasoning over your repos |
| Zero integration cost | No API build, no custom agent. Connect a repo and start scanning |
| Scheduled and targeted scans | Run on a cadence or trigger against a specific PR |
| Slack, Jira, and ticketing webhooks | Findings route into the workflows your security team already runs |
| Closed preview track record | Hundreds of orgs found vulnerabilities their SAST tools had missed for years |
| Coming next | Claude Team and Max plan access. Today, Enterprise tier only |
| Your real lever this week | Stand up scanning on one production repo, measure the gap against your existing SAST |
What Actually Shipped
Claude Security started life inside Anthropic as Claude Code Security, the research preview Anthropic announced in February 2026. The closed preview ran for roughly two months across hundreds of customer organizations. The public beta on April 30 is a renamed, productized version of the same thing, with three pieces that matter for an enterprise buyer.
It scans the way a security researcher scans, not the way a linter scans. Most static analysis tools match patterns. They look for known signatures and unsafe API calls and write a finding when the syntax matches. Claude Security reads source code and traces data flow. It reasons about how a value enters one file, passes through three layers of business logic, and lands inside a database query without proper escaping. That’s a reasoning task, not a pattern match. The Opus 4.7 model is what makes it feasible at codebase scale.
It writes fixes, not just findings. For every confirmed vulnerability, the platform produces a severity rating, a confidence score, an impact assessment, reproduction steps, and a targeted patch. The patch goes through human review. Your team approves or rejects. The work product is a pull request, not a ticket your devs have to triage from scratch.
It runs inside your Claude Enterprise tenant. The data boundary is the same data boundary your team already negotiated when they signed the Enterprise agreement. No new vendor review. No new DPA. No new SOC 2 report to chase. According to coverage in SiliconANGLE and SecurityWeek, the public beta added scheduled scans for ongoing coverage, the ability to dismiss findings with documented reasoning, and CSV plus Markdown exports for audit imports.
The integration story is the part that’s easy to miss. Claude Security ships with webhook integrations for Slack, Jira, and the ticketing systems most enterprise security teams already pipe their findings through. Your detection pipeline doesn’t change. The source of detections does.
Why “No Integration” Is the Whole Story
Most enterprise AI rollouts I see stall in the same place. The model is good. The use case is real. The procurement and integration work is what kills the timeline.
A typical AI security pilot inside a Fortune 500 looks like this. Three weeks for vendor security review. Two weeks for the data processing agreement red-line. Two weeks for the InfoSec architecture review. A month to stand up the API integration and the IAM scaffolding. Six weeks of pilot before anyone produces a real finding. Most pilots die between weeks four and ten because the integration cost outran the patience of the security team that asked for it.
Claude Security collapses that timeline because the procurement work happened when you signed the Claude Enterprise contract. The integration work is “connect a repo.” The data boundary review was already done. The same access controls and audit logging your team already trusts for Claude chat sessions cover Claude Security scans.
That’s why “no API integration required” is the whole story. The model quality matters less than most coverage gives it credit for. The reason this beta is interesting is that it removes the four to six weeks of plumbing that usually sits between an enterprise security team and a working AI security pilot.
The Closed Preview Result That Should Get Your Attention
The case-study claim from the closed preview is specific enough to take seriously. Hundreds of organizations participated. Many of them reported Claude finding vulnerabilities that their existing SAST tools had been running against the same code for years without surfacing.
This is the exact pattern Anthropic flagged in their zero-day research. When the Anthropic team pointed Claude Opus 4.6 at production open-source codebases earlier this year, it identified more than 500 vulnerabilities, including bugs that had survived years of expert review. The public beta runs on Opus 4.7, which is meaningfully stronger on multi-step reasoning than 4.6.
The reason existing SAST tools miss these bugs is structural. A pattern-matching scanner can find SQL injection in a single function. It struggles with a multi-file data flow where the unsafe value enters from a webhook, gets stored in a cache, gets read back two services later, and gets concatenated into a query inside a logging utility. That bug looks fine in any one file. The tool can only see one file at a time.
A reasoning model with a million-token context window can hold the entire code path in its head. That’s the categorical difference. It’s why Claude is finding bugs that SAST has missed. It’s why running Claude Security alongside your existing tooling, even for a single sprint, produces findings you’d have shipped to production otherwise.
What is Claude Security and how does it differ from SAST?
Claude Security is an AI-powered code scanner that uses Claude Opus 4.7 to read source code, trace data flows across files, and identify exploitable vulnerabilities. Unlike traditional SAST tools that match against known unsafe patterns, Claude Security reasons about how data moves through an application end-to-end, which lets it surface multi-file bugs that pattern matchers cannot detect.
Five practical differences from a typical SAST product:
- Reasoning over patterns. Traces data flow across files instead of matching known signatures.
- Patches with findings. Produces a targeted fix for human review, not just a finding ID.
- Codebase-scale context. Reads across files in a single reasoning pass.
- Confidence scoring. Each finding includes a severity rating and a confidence score, so triage prioritizes high-signal findings.
- Native enterprise data boundary. Runs inside the Claude Enterprise tenant your security team has already approved.
Where It Fits and Where It Doesn’t
A few honest qualifications that are worth saying out loud, because the beta-launch coverage is uniformly enthusiastic and the buying decision is more nuanced than that.
This is not a replacement for your existing security stack. Claude Security is complementary to SAST, DAST, SCA, and runtime security tooling. It finds a different class of bugs. It does not handle runtime monitoring, dependency vulnerabilities, infrastructure misconfiguration, or container image scanning. The right framing is “additional layer that catches what reasoning catches,” not “tool that replaces three line items in your security budget.”
The patch quality varies. Patches for narrow, single-function bugs are usually clean. Patches for architectural issues that require redesign are sometimes more like suggestions than fixes. Your team’s job is still to review every proposed change. The win is faster triage and a real starting point for the fix, not a hands-off “Claude wrote it, ship it” workflow.
Public beta means public beta. Pricing for general availability is not finalized. Capacity could throttle under heavy load. SLA commitments will land at GA, not now. If your team needs production-grade SLA today, you’re piloting, not switching production volume.
Access tier matters. As of April 30, the beta is gated to Claude Enterprise customers only. Anthropic has signaled that Claude Team and Max plan access is coming, but there is no public date. If you’re on Team or Max today, this beta isn’t for you yet.
The Strategic Read
Three things this launch tells you about where the frontier is moving.
Defensive cyber AI just stopped being a thought experiment. Two weeks ago, the conversation about AI in security was about GPT-5.4-Cyber under OpenAI’s verification gates and Claude Mythos under Project Glasswing partner agreements. Both are still gated. Claude Security is the first product where the frontier-grade reasoning model lands in your security team’s hands without an application process. That’s a different shape of release. The bar for “AI in production security workflows” just dropped from “you need a relationship with an AI lab” to “you need a Claude Enterprise contract.”
Anthropic is monetizing the model’s reasoning advantage as products, not just API tokens. The API-first business is selling tokens. The product business is selling outcomes. Claude Security is the second outcome-shaped product Anthropic has shipped in three weeks, after the Opus 4.7 design tool. The pattern says Anthropic is increasingly betting on first-party products tuned for specific high-value workflows, not generic API access. If you’re an enterprise buyer, that means more bundling, more outcome-aligned pricing, and more reason to run procurement at the contract level rather than the line-item level.
The partner ecosystem play is real. Anthropic’s coverage notes that CrowdStrike, Palo Alto Networks, SentinelOne, Trend Micro, and Wiz are integrating Opus 4.7 into their own platforms. That tells you the model is being licensed as a reasoning engine inside existing security tools, not just sold as a standalone product. The competitive pressure on traditional SAST and code review vendors is going to be material over the next two quarters.
The undercurrent worth tracking is the same one I covered in the Anthropic-Pentagon piece. Anthropic is leaning into commercial security products at the same time it’s locked out of direct DoD procurement. The commercial security market is not the federal market, but it’s not unrelated either, and the product velocity is what tells you where Anthropic’s commercial focus is.
What to Do This Week
Three concrete actions, all doable by Friday if you have a Claude Enterprise contract.
- Request public beta access through your Anthropic account team. The Enterprise admin in your Anthropic console should see the option. If not, your Anthropic CSM can enable it. Ask for usage credits to cover beta volume, and ask for a named technical contact for integration support during the beta. The first quarter of a public beta is when account teams have the most room to negotiate, and it’s when reference customer slots are most available.
- Pick one mid-stakes production repo and run a baseline scan. Don’t start with the crown jewels. Don’t start with a sandbox repo nobody cares about either. Pick something real but bounded. Run Claude Security against it. Run your existing SAST tool against the same repo on the same day. Compare the findings. The output of this exercise is a one-page table that shows what each tool caught and what only Claude caught. That table is the document that wins your CISO’s blessing for a wider rollout. Do this work yourself, not later, because the gap data is most credible when your own team produced it.
- Wire one webhook into your incident workflow. The Slack, Jira, or ticketing integration is the difference between “we ran a scan” and “we have a working pipeline.” Pick the one your security team already lives in. Route Claude Security findings into a triage queue. Run the queue for a sprint. Measure mean time to triage and rate of false positives. That measurement is what makes the GA pricing conversation tractable when it arrives.
If your org is on Team or Max plans rather than Enterprise, this isn’t the week. But it’s the right week to start the upgrade conversation. Anthropic has said Team and Max access is coming, and getting the procurement work in motion now means you’re ready when the gate opens.
Bottom Line
The interesting thing about Claude Security is that the technology was the easy part. Anthropic has been finding bugs in production code with Claude for months. The hard part was the productization, and they shipped it on April 30 with a procurement and integration model that removes the friction that usually kills enterprise security pilots.
You either treat this as a news item or a procurement event. Treating it as a news item is fine if your security backlog is empty and your existing SAST coverage feels complete. If neither of those is true, the move is to spin up a baseline scan this week, capture the gap data, and let the findings argue your case to the CISO.
The frontier security models that arrived in April are landing in three different shapes. GPT-5.4-Cyber is verification-gated. Claude Mythos is partner-gated. Claude Security is contract-gated, and the contract is one most enterprise buyers already have. The first-mover advantage on the third one is real, and the window where it’s still a competitive lever is narrow.
Do the scan. Pull the gap data. Send the memo. The customers who treat April 30 as a procurement event will be on their second sprint of triage by the time their peers are scheduling the vendor review.
Related Reading:
TAGS
Ready to Take Action?
Whether you're building AI skills or deploying AI systems, let's start your transformation today.
Related Articles
You're Measuring AI Adoption. Measure This Instead.
Gartner's 12,004-employee survey exposed the AI enablement illusion. Discover the proficiency metrics that actually predict enterprise AI ROI.
OpenAI on AWS Bedrock: Your Move This Week
OpenAI's GPT-5.5, Codex, and Managed Agents landed on AWS Bedrock April 28 with native IAM and PrivateLink. See the procurement move to make this week.
Custom GPTs Are Dead. Here's What Replaces Them.
OpenAI replaced Custom GPTs with Workspace Agents on Apr 22. See the migration framework before credit-based pricing kicks in on May 6.