Claude Security is now in public beta for Claude Enterprise customers. Anthropic says the tool scans codebases for vulnerabilities and proposes targeted patches for human review. For India and USA agency delivery teams, this matters because clients already ask for AI-assisted secure coding, faster triage, and clear audit trails.

This briefing sticks to verified details from Anthropic's product and blog posts. It focuses on what your delivery leads should check before they promise Claude Security inside a client engagement.

Claude Security public beta announcement visual from Anthropic — Feature image adapted from Anthropic's Claude Security public beta announcement. Credit: Anthropic / Claude brand assets.

Key Takeaways for Agency Delivery Teams

Claude Security is in public beta for Claude Enterprise; Team and Max access is coming soon.
Enterprise scans can run on Claude Mythos 5.1, with findings returned without giving users direct Mythos chat access.
GitHub.com and GitHub Enterprise Server repos are supported today; other hosts are not.
Every proposed patch still needs human review and approval before it lands.
Tech and services partners, including CrowdStrike, Palo Alto Networks, Accenture, Infosys, and PwC, are part of Anthropic's rollout path.
India and USA agencies should treat this as an Enterprise seat and GitHub App readiness check, not a plug-and-play add-on.

What Claude Security Is, in Plain Terms

Anthropic describes Claude Security as a capability built into Claude.ai. You pick a repository, start a scan, and get findings with severity, confidence, impact notes, and a recommended fix path.

The product was previously known as Claude Code Security. Anthropic says hundreds of organizations tested it in a limited research preview. Feedback from that period shaped the public beta.

You can open it from the Claude.ai sidebar or at claude.ai/security. Admins enable it in the admin console under Claude Security.

Anthropic frames the timing clearly: models are getting better at finding flaws, and the next wave will exploit them faster. Defenders need easier access to frontier scanning, not another custom agent build. See Anthropic's Claude Security public beta post for the company's stated rationale.

How a Claude Security Scan Works

A scan does not only hunt for known regex patterns. Anthropic says Claude reasons across files and modules, traces data flows, and reads source like a security researcher would.

From repository pick to findings list

Confirm Claude Enterprise, premium seats, Extra Usage, and the Claude GitHub App are in place.
Enable Claude Security in organization settings.
Open claude.ai/security and select a GitHub repository, branch, or directory.
Start the scan and wait; large repos can take minutes to hours.
Review findings, dismiss false positives with reasons, export, or open a remediation session in Claude Code on the Web.

Anthropic recommends scoping large repositories or monorepos to a directory. Narrower scope raises scan success and still lets the agent pull wider context when needed.

What each finding includes

Help Center docs list fields such as title, details, location, impact, reproduction steps, recommended fix, severity, status, category, repository, branch, and date created. Dismissed items keep a reason and optional note.

Severity is High, Medium, or Low. Anthropic assigns severity from exploitability in that codebase, not from the category label alone. The same injection class can land at different severities in different repos.

Claude Security product page visual showing Anthropic security tooling — Product visual from Anthropic's Claude Security page. Credit: Anthropic / assets.claude.com.

Claude Mythos 5.1, Opus 4.7, and Partner Paths

Anthropic's product and getting-started materials say Enterprise Claude Security scans run on Claude Mythos 5.1. Users get Mythos-quality findings without direct access to the model itself.

Interactive patching opens in Claude Code on the Web and uses the models already available in your account. The Mythos scan surface does not extend Mythos chat to other Claude screens.

The public beta announcement also highlights Claude Opus 4.7 for vulnerability work and for partner integrations. Treat Mythos as the Enterprise scan engine documented for claude.ai/security, and Opus 4.7 as the broadly available model Anthropic is pushing into partner security tools.

In a related update, Anthropic said it is integrating Mythos-class defense into partner products and launching a $35 million Defender Advantage Fund in Claude credits for open-source security work.

Named technology and services partners

Technology partners named in the public beta post include CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, and Wiz. Services partners include Accenture, BCG, Deloitte, Infosys, and PwC.

For agencies, that means three adoption paths: direct Claude Security, a security platform that embeds Opus or Mythos outcomes, or a services-led rollout. Pick the path that matches the client's stack, not the shiniest demo.

What Claude Security Looks For

Anthropic's help docs group example finding types into buckets agencies already know from secure SDLC work:

Injection - SQL, command, code, XSS, XXE, and ReDoS style issues.
Path and network - path traversal, SSRF, and open redirects.
Auth and access - auth bypass, privilege escalation, IDOR/BOLA, CSRF, and race conditions.
Memory safety - buffer or integer overflow, use-after-free, unsafe misuse, mostly in C/C++/Rust unsafe code.
Cryptography - timing leaks, algorithm confusion, and weak primitives.
Deserialization and protocol issues - unsafe object construction and encoding confusion.

The product page says Claude Security focuses on high-severity classes such as memory corruption, injection, authentication bypasses, and complex logic errors that pattern tools often miss.

Scans are stochastic by design. Anthropic notes that unlike fixed static analyzers, the agent adapts each run. That depth helps catch logic bugs, but agencies should expect some result variance across repeats.

Setup Checklist Before You Pitch a Client

Do not sell Claude Security as a weekend toggle. Anthropic's getting-started guide lists hard prerequisites.

Claude Enterprise account - public beta is Enterprise-only right now.
Premium seats - each scanner needs a premium seat; standard seats do not include Claude Code on the Web.
Claude Code on the Web enabled - remediation sessions depend on it.
Extra Usage enabled - scans use consumption billing; set spend limits.
Claude GitHub App installed - with access to the target repos; GHES must already connect to Claude Code on the Web.
Admin enablement - turn on Claude Security in organization settings.

Billing note from Anthropic: scans are charged at direct token cost with no separate Claude Security platform fee. Mythos scans are billed as standard token usage under the existing plan.

Scope of use is strict. Anthropic says you may only scan code your company owns and has rights to scan. Do not point it at random third-party open-source repos outside your client's codebase.

Delivery Playbook for India and USA Agency Teams

India and USA delivery teams often share the same client brands but different repo layouts, compliance pressure, and review bandwidth. Use one shared playbook with local owners.

Week-one discovery questions

Is the client already on Claude Enterprise with premium seats?
Are repos on GitHub.com or GitHub Enterprise Server only?
Who owns dismissals, patch review, and merge rights?
Where do findings need to land: Jira, Slack, CSV audit folders, or a GRC tool?
Which directories are in scope for the first Extended scan?

Operating cadence that survives real sprints

Anthropic suggests scheduled scans and names a weekly cadence as a common fit. Tie the schedule to a Monday triage, a sprint boundary, or a pre-release gate. Give findings a named owner or the backlog will rot.

Use webhooks for scan-complete and new-finding events. Export CSV or Markdown for audit packs. Document dismissal reasons so the next reviewer trusts prior decisions.

If your client is still designing the broader secure-delivery model, pair this tool briefing with Debugged Pro's AI security, risk, and compliance services and the recent OpenAI rogue AI agents agency guide. Those pages help frame governance while Claude Security handles repo-level discovery.

India-specific delivery notes

Many India delivery pods own shared services for US product orgs. Confirm GHES IP allow lists, SSO authorization for the Claude GitHub App, and which India leads hold premium seats. Keep patch approval with the code owners named in the US client RACI, even when India runs the scan queue.

USA-specific delivery notes

US enterprise buyers will ask about data retention, ZDR constraints, and whether Mythos access is exposed. Anthropic states Mythos is used for scans without exposing the model for free-form prompting. Still put human patch review in the statement of work.

For broader build-and-consult packaging, see Debugged Pro's software development consulting page and keep Claude Security as one control in the delivery plan, not the whole security story.

Where Claude Security Helps Agencies, and Where It Does Not

Helpful when: the client already pays for Claude Enterprise, code lives on supported GitHub hosts, and security and engineering can meet in one remediation loop.

Weak fit when: repos sit on GitLab or Bitbucket only, the org lacks premium seats, or leadership expects zero human review. Anthropic is clear that people must approve patches.

Also remember scan consistency is not identical run to run. Budget a second pass on critical modules instead of treating one scan as a certificate.

Traditional SAST still has a place for known pattern coverage and CI gating. Claude Security is closer to a researcher-style pass that explains context and proposes a fix path.

Practical Pilot Plan You Can Sell This Month

Pick one service repo with clear ownership and recent change volume.
Run a Standard scan, then an Extended scan on the hottest directory.
Triage High findings in one sitting with security and the module owner.
Open remediation in Claude Code on the Web for two or three accepted issues.
Export the paper trail and wire Slack or Jira webhooks before you expand scope.
Write the RACI for India scanners and USA approvers before the second repo.

Measure time from scan complete to reviewed pull request, not ticket count. Anthropic's early-user quotes stress minutes-to-PR as the signal that matters.

Risk and Compliance Notes for Client Conversations

Always review proposed patches before applying them, especially on critical systems. Anthropic says Claude can make mistakes.

Help Center notes that Anthropic may retain data where required by law or to address Usage Policy violations, and calls out No Zero Data Retention constraints for this product surface. Confirm that wording against the client's legal review before you put it in a proposal.

If a client's security team needs reduced cyber safeguards for authorized defensive work on Opus or Sonnet, Anthropic points them to the Cyber Verification Program. That is separate from Claude Security scan access.

What Delivery Managers Should Tell Stakeholders Tomorrow

Claude Security is real, Enterprise-gated, and GitHub-centered today.
Mythos 5.1 powers Enterprise scans on the Security app; patch sessions use account models.
Partner platforms and services firms are alternate entry points if the client will not buy Claude seats yet.
Human review stays non-negotiable in every statement of work.
India and USA pods can share one scan cadence if ownership and merge rights are explicit.

If nothing else, update your discovery checklist this week. Ask every AI-curious enterprise client whether Claude Enterprise and GitHub App access already exist. That single answer decides whether Claude Security is a near-term delivery lever or a roadmap item.

FAQ: Claude Security for Agency Teams

Who can use Claude Security today?

Anthropic says Claude Security is in public beta for Claude Enterprise customers. Access for Claude Team and Max customers is coming soon. Admins enable it in the admin console.

Which repositories can Claude Security scan?

GitHub.com and GitHub Enterprise Server repositories are supported. Anthropic says other hosting providers are not supported today. You should only scan code your organization owns and has rights to scan.

Does Claude Security use Claude Mythos?

Anthropic's product and getting-started docs say Enterprise scans on claude.ai/security run on Claude Mythos 5.1. Users receive findings without direct Mythos model access. Patching in Claude Code uses the models available in your account.

How are Claude Security scans billed?

Anthropic states scans are charged at direct token cost with no additional Claude Security platform fee. Mythos scans are billed as standard token usage under the existing Enterprise plan. Extra Usage must be enabled.

Can agencies auto-merge Claude Security patches?

No. Anthropic requires human review and approval before patches are implemented. Agencies should keep merge authority with named code owners in the RACI for India and USA delivery teams.

About the Author

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA delivery teams on AI product rollouts, secure engineering practices, and client-ready operating cadences.