What is Claude Security? Anthropic’s AI Vulnerability Scanner (May 2026)

Claude Security is Anthropic’s AI-powered defensive security tool — a vulnerability scanner that uses Claude Opus 4.7 to find, validate, and fix code security flaws semantically rather than via signature matching. It entered public beta on April 30, 2026 and represents Anthropic’s first commercial cybersecurity product.

Last verified: May 6, 2026

The 60-second summary

• What: AI-powered security scanner running in Claude Code on the web.
• Model: Claude Opus 4.7.
• Released: Public beta, April 30, 2026 (after closed preview since late 2025).
• Pricing: No platform fee — pay direct Opus 4.7 token costs only.
• Access: Claude Enterprise customers at launch; Team and Max coming soon.
• Differentiator vs Snyk/GHAS: Semantic data-flow tracing + self-validation, not signature matching.
• Related projects: Project Glasswing (Mythos-based, critical infrastructure focus), Claude Mythos Preview (research-only).

What Claude Security actually does

Three workflows ship in the May 2026 public beta:

1. Codebase scanning

Point Claude Security at a repository in Claude Code on the web. It:

• Reads the codebase structure and identifies entry points (HTTP handlers, file ingestion, message queue consumers).
• Traces data flow from each entry point through transformations, validations, and sinks.
• Identifies common vulnerability classes: SQL injection, XSS, SSRF, path traversal, command injection, insecure deserialization, broken authentication, race conditions, secret exposure.
• Self-validates each candidate finding by re-running its own check with adversarial reasoning (“could this actually be exploited?”).
• Surfaces only validated findings with severity scoring and exploit explanation.

Scan time: minutes for small repos, 30-90 minutes for large monorepos. Token cost: ~$5-25 per scan typically.

2. Targeted remediation

For each finding, Claude Security generates:

• A suggested fix as a code diff.
• Justification (“this fix closes the vulnerability because…”).
• Test cases to verify the fix.
• Risk-of-regression analysis (“this change might also affect X, Y”).

Fixes are presented as suggestions, not auto-applied. The reviewer decides whether to commit.

3. CI/CD integration

Claude Security can be invoked from CI pipelines — scan a PR diff, comment with findings, block merge on critical issues. Integration is via Claude Code’s existing GitHub/GitLab integrations rather than a dedicated security CI app.

How Claude Security differs from rule-based scanners

The fundamental architectural difference, in three points:

Capability	Rule-based (Snyk, GHAS, Semgrep)	Claude Security
Detection method	Pattern match against rule library	Semantic data-flow tracing
Scan speed	Seconds	Minutes to hour+
Cost per scan	Flat per-seat / per-repo	$5-25 token cost per scan
False positive rate	High (15-40% commonly reported)	Low (Anthropic claims <5%)
Novel vulnerability detection	Poor — needs new rule	Good — reasons about new patterns
Coverage of known CVEs	Excellent — direct signature match	Good — but slower and per-token
Best for	High-frequency CI scanning, dependency audit	Pre-release deep audit, novel app review

In practice, mature security teams will likely run both: rule-based scanners on every commit for known-CVE coverage, Claude Security on PRs touching security-sensitive code or pre-release audits.

The self-validation design choice

The most discussed technical detail in May 2026 industry coverage is the self-validation step. Most LLM-based scanners flag candidate vulnerabilities with high false positive rates. Claude Security adds a second pass:

1. First pass: identify candidate vulnerability.
2. Second pass: “Try to construct an actual exploit for this finding. Can you?”
3. Surface decision: if step 2 produces a working exploit, surface as confirmed; if not, drop the finding.

This roughly halves the false positive rate at the cost of doubling token spend per finding. For security teams, that tradeoff is highly favorable — false positives are the #1 reason security tooling gets ignored by developers.

Pricing and access in May 2026

Pricing structure:

• No platform fee for Claude Security itself.
• Pay direct Claude Opus 4.7 API token rates: $15/1M input, $75/1M output.
• A typical scan: $5-25 in token cost.
• Heavy monorepo or deep audit: $50-200+.

Access tiers:

• Claude Enterprise: available now (May 2026).
• Claude Team and Max: “coming soon” per Anthropic.
• Pro and Free: no announced timeline.

Comparable enterprise tooling:

• Snyk Enterprise: $98+/developer/month, plus per-repo costs.
• GitHub Advanced Security: $49/committer/month on top of GitHub Enterprise.
• Checkmarx / Veracode: custom enterprise pricing, typically $50-200K+/year.

For a 100-engineer org running Claude Security on every PR, expected monthly cost is $3-10K — competitive with mid-tier rule-based scanner subscriptions, with the offsetting benefit of fewer false positives.

Limitations and open questions

Five caveats from May 2026 industry coverage and Reddit reports:

1. Enterprise-only at launch. Smaller teams and individual developers locked out until Team / Max rollout.
2. Speed. Slow compared to rule-based scanners; not suited for tight inner-loop dev experience.
3. Cost variance. Token-based pricing means a malformed scan command can run up unexpected bills.
4. Human review still required. Anthropic explicitly designs in approval gates — fixes aren’t auto-applied. Some buyers reading this as “still needs a human babysitter” rather than mature autonomy.
5. Coverage of dependency vulnerabilities. Claude Security focuses on first-party code. Dependency-CVE scanning still belongs to Snyk-class tools.

How Claude Security relates to Glasswing and Mythos

Three Anthropic security initiatives, three different scopes:

Product	Model	Audience	Status
Claude Mythos Preview	Mythos	Anthropic researchers, curated partners	Not commercial
Project Glasswing	Mythos + Opus 4.7	Critical OSS maintainers (Linux, browsers)	Partner program
Claude Security	Opus 4.7 only	Enterprise developers	Public beta

Mythos has been used internally to find thousands of zero-day vulnerabilities — the Anthropic CEO Dario Amodei publicly described this as “AI’s moment of cyber danger” in a CNBC interview on May 5, 2026. Glasswing is the responsible disclosure and remediation effort partnering with critical OSS maintainers. Claude Security is the productized, broadly available defensive scanner anyone can pay for — running on the older but commercially available Opus 4.7 model.

Bottom line

In May 2026, Claude Security is the first credible LLM-native vulnerability scanner with enterprise-grade design — semantic data-flow tracing, self-validation to cut false positives, and transparent pay-per-token pricing without platform fees. It complements rather than replaces rule-based scanners like Snyk and GitHub Advanced Security, with a clear role in pre-release deep audits and PR-level review of security-sensitive code. For Claude Enterprise customers, it’s the obvious tool to evaluate immediately. For everyone else, the Team / Max rollout will determine accessibility through 2026.

Sources: Anthropic Help Center (May 2026), Help Net Security coverage (May 4, 2026), The New Stack coverage (April 30, 2026), SiliconANGLE coverage (April 30, 2026), CNBC Amodei interview (May 5, 2026), ZDNET coverage (May 2026).