1300 166 138 Tell us what you need
News / Security

Anthropic unveils Claude Security for code scans

By Monika Tantau • Published 7 months ago
Anthropic unveils Claude Security AI vulnerability scanner for codebases

Key takeaways

• Anthropic launched Claude Security in public beta for Enterprise-tier customers, with Team and Max access coming soon
• The tool scans full repos or targeted directories, traces data flows, and runs a multi-stage validation pipeline with confidence ratings
• Workflow features include scheduled scans, dismissible findings with documented reasons, and CSV/Markdown export for audit trails
• Anthropic tightened Opus 4.7 cyber safeguards and gated offensive capabilities behind a Cyber Verification Program for vetted researchers

Most Read

Anthropic has released Claude Security, an AI-driven scanner that reads full codebases, traces data flows, and produces ranked vulnerability findings with confidence ratings. The tool is in public beta for Enterprise-tier Claude customers, with Team and Max-tier access flagged as coming soon.

The launch, covered by ZDNet, lands in a week dominated by AI security news on both sides of the table. For Australian SDLC teams running regulated workloads under ASIC, APRA, and OAIC requirements, the question is no longer whether AI-driven SAST belongs in the stack. It is which vendor and which workflow.

What Claude Security actually scans

Claude Security operates on full repositories or targeted directories. Rather than running pattern-matching across files in isolation, it reasons about source code as a connected system. The model traces data flows, follows function calls across modules, and reports how components interact. Each finding includes severity, likely impact, reproduction steps, and a recommended fix.

That cross-file reasoning is the practical difference from regex-based static scanners. Traditional tools flag suspicious patterns and leave engineers to confirm whether the path is exploitable. Claude Security produces context-aware findings tied to specific code paths, which shifts triage time from confirmation toward remediation. The Enterprise-tier scope at launch indicates Anthropic is targeting organisations with mature security operations rather than individual developers.

Confidence ratings change the triage workflow

Each Claude Security finding passes through a multi-stage validation pipeline that independently checks the result and assigns a confidence rating. Infosecurity Magazine noted that the design directly targets the false-positive fatigue that has plagued enterprise SAST adoption for years.

Workflow features round out the offering. Scheduled scans provide ongoing repository coverage. Findings can be dismissed with documented reasons, creating an audit trail. CSV and Markdown export connect results to existing ticketing and audit systems. For teams already running mature document workflow across compliance and risk functions, the export hooks make Claude Security a candidate to plug into established remediation processes rather than a parallel tool to maintain.

Why Australian SDLC teams should evaluate it now

Australian regulated industries have spent the past two years tightening notifiable data breach response under the OAIC scheme, while APRA’s CPS 234 standard has pushed banks and superannuation funds to demonstrate active vulnerability management. The combination has made fast, credible triage of code-level risk a board-visible metric in financial services and health.

AI-driven SAST changes the economics of meeting those obligations. A scanner that reduces false positives and produces audit-ready exports lets a smaller team cover more code without losing traceability. For Australian shops with mature DevSecOps practices, the integration cost is modest relative to the gain. For shops without those practices, Claude Security is a forcing function. Either way, supply-chain exposure through public package registries means doing nothing is no longer the cheapest option.

Anthropic Opus 4.7 cyber safeguards address dual-use risk of AI vulnerability scanners

The dual-use risk Anthropic tried to hedge

A scanner that can find chained exploits is also useful to attackers, and Anthropic has acknowledged that. Opus 4.7 ships with expanded cyber safeguards designed to block prompts that match prohibited or high-risk offensive cybersecurity uses. Approved researchers can request restricted access through a Cyber Verification Program, and Anthropic has separately disclosed Project Glasswing alongside a restricted internal model called Mythos that is not being released publicly.

The recent Mythos disclosure showed why the gating matters. The same reasoning capability that lets Claude Security identify vulnerabilities for defenders can locate and chain zero-days for attackers. Anthropic’s bet is that customer-facing defensive products plus tightly scoped researcher access is the right balance. Whether that holds will depend on how visible misuse becomes once Team and Max-tier access opens.

A new layer in the AU security stack

AI-driven SAST is not replacing traditional scanners, bug bounty programs, or penetration testing. It joins them as a fourth layer, sitting closer to the development cycle and producing higher-confidence findings at lower marginal cost.

For Australian organisations the practical question is sequencing. The teams that integrate this layer early, refine the workflow, and feed export data into existing audit pipelines will set the benchmark that regulators and boards eventually treat as the expected standard. The rest will be measured against it.

Scroll to Top