Claude Security opens public beta with Opus 4.7 repo scans

• Anthropic opened Claude Security in public beta, and claudeai's launch post says it scans repos for vulnerabilities, validates findings to cut false positives, and drafts patches for human review.
• The product is packaged as an on-ramp for teams that want Opus 4.7 on production code without building their own agent stack, according to claudeai's thread reply and Anthropic's Getting started guide.
• The beta adds scheduled scans, directory-level targeting, CSV and Markdown exports, webhook notifications, and dismissals that persist across future scans, per claudeai's feature list and the Claude Security help doc.
• Anthropic's docs put the product inside Claude Code on the web, where repos run in a remote GitHub-backed environment, as _catwu's post and the Claude Code on the web guide both indicate.
• The quiet gotcha is access: launch tweets frame Claude Security as Enterprise-only, while Anthropic's updated help center article says public beta access now covers Max, Team, and Enterprise.

You can read the product page, the full getting started guide, and the updated help center doc. There is also an older research preview screenshot showing Security Scan already sitting in Claude Code's sidebar, plus rohanpaul_ai's rundown surfacing the branch, directory, effort-mode, and webhook knobs that matter in practice.

The core loop is simple: pick a GitHub repo, run a scan, inspect validated findings, then jump straight into a remediation session. Anthropic's product page says the model reasons across files, traces data flows, and runs an adversarial verification pass before it shows a result.

That puts the product closer to a built-in code security workflow than a generic chat wrapper. claudeai's thread reply explicitly pitches it as the way to use Opus 4.7 on code without standing up a custom API integration or agent build.

Anthropic's getting started guide says the feature depends on Claude Code on the web, extra usage billing, the Anthropic GitHub App, and premium seats for users who run scans. The same guide says admins can set a separate spend limit for Claude Security after it is enabled.

The current repo boundary is narrow. The help center doc says scans only work on GitHub.com repositories, while the getting started guide recommends scoping large repos or monorepos to a directory to improve reliability and determinism.

Anthropic also exposes a few execution knobs that matter for real usage:

• Branch and directory scoping, per rohanpaul_ai's feature summary and the getting started guide
• Regular versus Extended effort, with Extended positioned as the deeper pass in the getting started guide
• Parallel projects and recurring schedules, per claudeai's launch thread and the help center doc

Anthropic says hundreds of organizations used the research preview on production code before this beta, and claudeai's thread claims those teams found issues that existing scanners had missed. The product page frames the pitch the same way: fewer noisy alerts, more validated findings, and suggested fixes that preserve code structure and style.

The help docs are unusually concrete about what a finding contains. Each result includes title, details, file path and line number, impact, reproduction steps, recommended fix, severity, status, category, repository, branch, and created date, according to the Claude Security help article.

The vulnerable code categories span more than the usual web-app checklist. Anthropic lists injection, XXE and ReDoS, path traversal and SSRF, auth bypass and IDOR, memory safety issues, crypto mistakes, deserialization bugs, and protocol or encoding confusion in the help doc.

Once a finding lands, teams can:

• export a project, scan, or finding to CSV or Markdown, per the getting started guide
• push scan-completion or new-finding events to Slack, Jira, or other systems through per-project webhooks, per the same guide
• open a remediation session in Claude Code on the web to generate a candidate patch, as claudeai's launch post and rohanpaul_ai's summary both show
• dismiss false positives with a reason that carries forward across future scans, per claudeai's thread and the help center doc

The launch messaging is narrower than the docs. claudeai's availability post says Claude Security is available in public beta for Claude Enterprise customers, and the product page uses the same framing.

But Anthropic's updated help center article now says, in plain text, that Claude Security is available in public beta for Max, Team, and Enterprise users. That lines up with testingcatalog's post, which speculated Team access would show up, and it is the clearest sign in the source set that Anthropic widened access faster than the launch thread suggested.