Editorial illustration for Claude Security Runs Inside Claude Code as Multi-Agent Vulnerability Scanner
Claude Security Scans Code Vulnerabilities Built-In
Claude Security Runs Inside Claude Code as Multi-Agent Vulnerability Scanner
Anthropic shipped Claude Security as a built-in feature of Claude Code this year, skipping the usual step of making customers wire up a separate API key. OpenAI followed with Codex Security, releasing its CLI under an Apache-2.0 license, though actually running a scan still requires access to OpenAI's models. Cloudflare went further on openness, publishing its scanning skill under MIT and designing it to work with any model, a bet that's paid off with 47,500 GitHub stars, more than any competitor. Google rounded out the field with Mantis, a tool built to reproduce vulnerabilities inside a sandbox, with Google's own team warning that it should be isolated when deployed.
Four labs, four licensing choices, four sets of assumptions about who gets to use these tools and how much trust they deserve. But strip away the branding and the access models, and the underlying architecture is nearly identical across all of them. Each company arrived at the same answer to a problem that has dogged AI-assisted security review since the first LLM-based scanners appeared: how do you stop a bug-hunting agent from drowning its own output in false positives.
The key finding: all 4 tools converge on one design. One agent finds a bug, a separate agent tries to disprove it, and only survivors get a patch. The differences are in where they run, who can use them and how much proof they demand.
Why this matters
Four separate companies landing on the same find-disprove-patch architecture within months of each other tells us this is less a product category than a default pattern, the way REST APIs or CI pipelines became defaults. For developers, that's good news: adversarial verification between agents is a real check against the false-positive noise that's plagued automated scanning for years. But the convergence also means the real competition isn't the architecture, it's deployment.
Claude Security running locally inside Claude Code versus Cloudflare's or Google's hosted versions changes who sees your code, who controls the threat model, and who gets to decide what counts as "proof" a bug is real. Founders evaluating these tools should ask less "does it find bugs" and more "where does my source go, and whose disprove-agent am I trusting." Researchers should watch whether this adversarial-pair pattern generalizes beyond security scanning into other agent workflows, since four labs agreeing on a design usually means someone else is already building the fifth copy.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
From Archives Professor Berinsky: AI chatbots can help refute misinformation
From Archives Google’s Gemini 4 Carbon Tops Charts Again, but Only for Now
From Archives Anthropic IPO: USD 2.08 Revenue Per USD 1 Compute Highlights Infrastructure Exposure
Business & Startups
From Archives OpenAI doubles Pro usage limits after pushback as Anthropic competition heats
Research & Benchmarks
Common Questions Answered
How does Claude Security differ from OpenAI's Codex Security in terms of deployment?
Claude Security is built directly into Claude Code as an integrated feature, eliminating the need for customers to configure separate API keys. In contrast, OpenAI's Codex Security requires users to access OpenAI's models to run scans, even though the CLI tool itself is released under an Apache-2.0 license for broader accessibility.
What is the find-disprove-patch architecture used by vulnerability scanners?
The find-disprove-patch architecture is a multi-agent design where one agent identifies potential bugs, a separate agent attempts to disprove those findings, and only the vulnerabilities that survive this adversarial verification process receive patches. This approach significantly reduces false positives that have historically plagued automated security scanning tools.
Why did Cloudflare's scanning tool achieve more GitHub stars than competitors?
Cloudflare published its scanning skill under the MIT license and designed it to work with any model, prioritizing openness over vendor lock-in. This approach resonated with the developer community, resulting in 47,500 GitHub stars, more than any competing vulnerability scanner tool.
What does the convergence on the find-disprove-patch design pattern indicate about the security scanning market?
The fact that four separate companies independently adopted the same find-disprove-patch architecture within months suggests this is becoming a default pattern rather than a differentiating product feature, similar to how REST APIs and CI pipelines became industry standards. This means real competition will shift from architectural differences to deployment flexibility and accessibility.
Further Reading
- Making frontier cybersecurity capabilities available to defenders - Anthropic
- Claude Security is now in public beta - Claude by Anthropic
- Scan your codebase for vulnerabilities - Claude Code Docs
- Alberta uses Claude to find and fix security vulnerabilities - Anthropic
- How we contain Claude across products - Anthropic
