TL;DR: Anthropic opened Claude Security to public beta for Claude Enterprise customers — a tool that scans codebases, validates findings with an adversarial verification pass, and proposes targeted patches for human review. A companion Claude Security Plugin for Claude Code is also in beta, scanning directly inside your coding-agent environment. Anthropic says it's built on the same models Anthropic uses to secure its own codebase.
What Anthropic launched
Anthropic's product page frames Claude Security around a three-step workflow: scan, validate, patch.
"Claude scans your codebase, validates findings, and suggests patches you can review and approve. Available in public beta for Claude Enterprise."
The pitch draws an explicit comparison to human expertise:
"Claude reasons through your code like a skilled security researcher. It understands context, traces data flows, and catches vulnerabilities that pattern-matching tools miss. Then it suggests a fix."
That "pattern-matching tools miss" line is a direct jab at how a lot of traditional static analysis works — matching known vulnerability signatures against code syntax, which catches known issue classes but tends to miss vulnerabilities that only become apparent when you trace how data actually flows across multiple files and functions.
TL;DR — Claude Security at a glance
| Question | Answer |
|---|---|
| What is it? | Codebase security scanner with validation and patch suggestions |
| Status? | Public beta |
| Who can use it? | Claude Enterprise customers |
| Access model? | Contact sales — not self-serve |
| Companion plugin? | Claude Security Plugin for Claude Code, also in beta |
| Auto-fixes code? | No — every patch requires human review and approval |
| Integrates with? | Slack and Jira via webhooks; export for audit |
| Can you scope scans? | Yes — partial repo scans and scheduled recurring scans |
| Model basis? | Anthropic's most capable generally available models |
The three-stage pipeline
Anthropic describes the workflow in three explicit stages:
1. Scan your code in parallel
"Claude Security understands context, traces data flows across files, and identifies complex, multi-component vulnerability patterns that traditional scanners might not detect."
The emphasis on cross-file data-flow tracing matters technically: a lot of real-world vulnerabilities (like injection flaws or improper authorization checks) only become visible when you follow a variable from an untrusted input, through several function calls and file boundaries, to where it's used unsafely. Signature-based scanners that look at code in isolated chunks tend to miss this class of issue, while an LLM with a large enough context window and the ability to actively navigate a codebase can, in principle, follow that chain the way a human security researcher would during manual review.
2. Validate findings
"Every finding goes through an adversarial verification pass. Claude challenges its own results before surfacing them. More real issues get reported, and fewer false positives waste analyst time."
This is the most interesting design choice, and it addresses a well-known pain point in application security: alert fatigue from false positives. Traditional SAST (static application security testing) tools are notorious for generating large volumes of findings that require manual triage, and security teams often deprioritize or ignore scanner output as a result. An adversarial self-check — where the model attempts to disprove its own initial finding before reporting it — is a reasonable mechanism for cutting noise, though Anthropic hasn't published a false-positive rate comparison against specific competing tools alongside this announcement.
3. Review and patch
"Claude detects issues and proposes fixes. Every finding includes a recommended patch for teams to review and approve. Fix vulnerabilities quickly rather than adding them to a growing backlog."
Anthropic frames patches as targeted — maintaining the code's existing structure and style rather than large, sweeping rewrites, which matters for review speed. A patch that looks like it belongs in the existing codebase is faster for a human reviewer to evaluate and trust than a patch that reformats or restructures unrelated code alongside the fix.
Claude Security Plugin for Claude Code
Alongside the hosted product, Anthropic shipped a Claude Security Plugin for Claude Code, also in beta:
"Scan, validate, and patch directly inside Claude Code while keeping all code inside your environment."
This targets a specific concern: teams that don't want code leaving their own infrastructure to reach a separate hosted scanning service. Running security scanning as a Claude Code plugin keeps the workflow inside the same coding-agent session developers already use, rather than requiring a context-switch to a standalone security dashboard — a meaningful adoption-friction reduction if teams are already using Claude Code for development.
How it fits into existing security workflows
Anthropic lists several integration points aimed at security teams specifically, not just individual developers:
| Feature | What it does |
|---|---|
| Transparency and visibility | Every finding includes the vulnerability, why it matters, and the proposed fix |
| Targeted patches | Fixes preserve code structure and style |
| Webhooks to Slack/Jira | Findings route into existing team communication and ticketing |
| Audit export | Findings can be exported for compliance/audit records |
| Scheduled recurring scans | Not a one-time run — can be set to repeat |
| Partial repo scans | Scope scans to specific parts of a codebase, not just full-repo runs |
| Full remediation control | Every patch requires human review and approval |
The human-review requirement is worth underlining given the broader conversation about autonomous AI agents taking action without oversight. Claude Security explicitly stops short of auto-merging fixes — it's positioned as an assistant that accelerates the triage-and-patch cycle, not an autonomous remediation system that acts independently on a codebase.
Context: Anthropic's own security research
Claude Security doesn't launch in isolation — it sits alongside Anthropic's ongoing public research on AI and offensive/defensive cybersecurity, including work on Claude Mythos and Project Glasswing, Anthropic's cyber-defense research initiative, and public posts on evaluating and mitigating the risk of LLM-discovered zero-days. Anthropic's own framing — "the same models Anthropic uses to secure its own codebase" — ties Claude Security directly to that broader research program rather than treating it as a bolt-on product feature unrelated to the company's security research.
This matters for how to read the product: Claude Security isn't just a repackaged code-review feature. It's a productized version of a capability Anthropic has been separately researching and publicly discussing as both an opportunity (AI-assisted vulnerability discovery for defenders) and a risk (the same capability could theoretically assist attackers if capability controls aren't in place) — a tension that has shaped Anthropic's export-control and access-restriction decisions on frontier models more broadly.
What this means for security and platform teams
If you're on Claude Enterprise:
- This is a contact-sales feature, not self-serve — reach out to Anthropic directly to get access to the beta.
- Plan for a pilot on a subset of your codebase first, using the partial repo scan capability, before rolling scans across your full estate.
- Wire up the Slack/Jira webhook integration early so findings land in your team's existing triage workflow rather than a separate dashboard nobody checks.
If you're not on Enterprise:
- The Claude Code plugin is the more accessible on-ramp if your team already uses Claude Code for development — check current plugin availability and tier requirements directly with Anthropic, since exact plugin-tier gating wasn't detailed in the public announcement.
- Compare against your existing SAST/DAST tooling on false-positive rate specifically — that's the metric Anthropic is most directly claiming an advantage on, and it's the easiest one to validate yourself on a real codebase during a pilot.
Update — July 24, 2026: Cisco’s Antares open-weights a different defender lane — local SLMs that rank vulnerable files from CWE/advisory context (VLoc Bench).
Related on explainx.ai
- Cisco Antares — open-weight vulnerability localization (Jul 2026)
- Hugging Face was breached by OpenAI's own models during a cyber eval — the defender-guardrail asymmetry Claude Security is a partial answer to
- Claude Mythos preview — cybersecurity and Project Glasswing
- AI cyber guardrails debate — Kimi K3 and GLM 5.2 block US defenders
- Claude Code commands — complete reference guide
- US export controls on Fable 5 and Mythos 5 — the full story
Primary sources: claude.com/security · Anthropic blog: "Evaluating and mitigating the growing risk of LLM-discovered 0-days" · Anthropic blog: "Experimenting with AI to defend critical infrastructure"
Product details reflect Anthropic's Claude Security public beta announcement as of publication. Beta features, pricing, and availability change — verify current status and enterprise access requirements directly with Anthropic before planning a rollout.