OpenAI launched Codex Security on March 6, entering the AI-powered vulnerability scanning market fourteen days after Anthropic released Claude Code...
The Signal
OpenAI launched Codex Security on March 6, entering the AI-powered vulnerability scanning market fourteen days after Anthropic released Claude Code Security on February 20. Both tools use LLM reasoning instead of pattern matching, and both independently identified vulnerability classes that traditional SAST tools are structurally incapable of detecting. Combined, the two labs carry a private-market valuation exceeding $1.1 trillion. Both scanners are currently free to enterprise customers in research preview. The scale of findings is substantial. Anthropic's Claude Opus 4.6, released February 5, surfaced more than 500 previously unknown high-severity vulnerabilities in production open-source codebases, including a heap buffer overflow in the CGIF library that coverage-guided fuzzing missed despite 100% code coverage. OpenAI's Codex Security, evolved from an internal GPT-5-powered tool called Aardvark, scanned more than 1.2 million commits during beta, producing 792 critical findings, 10,561 high-severity findings, and 14 assigned CVEs across projects including OpenSSH, PHP, and Chromium. False positive rates fell more than 50% during beta; over-reported severity dropped more than 90%. Independent validation introduced caveats. Checkmarx Zero researchers found that Claude Code Security identified eight vulnerabilities in a full production codebase scan, but only two were true positives, with moderately complex obfuscation defeating detection. Neither lab has submitted its detection claims to independent third-party audit.
Why It Matters
The real story here isn't the detection capability—it's the deliberate commoditization of SAST as a strategic land-grab. OpenAI and Anthropic aren't giving away scanners out of altruism; they're repricing an entire category to zero in order to own the trust relationship with enterprise security teams before their IPO roadshows begin. Free today means switching costs and platform lock-in tomorrow, and security leaders who treat this as a procurement windfall are missing the commercial architecture being built around them. What surface coverage consistently undersells is the Checkmarx validation finding: a 75% false positive rate in production conditions isn't a beta quirk, it's a signal that the headline numbers from both labs were generated in controlled conditions that don't survive contact with real codebases. The 792 critical findings and 500-plus zero-days are genuine and alarming, but they were produced at a scanning scale and codebase selection that flatters the technology. When you run these tools against your estate with its obfuscated legacy modules and third-party integrations, the performance envelope shrinks considerably. The deeper pattern to watch is what happens when two IPO-motivated labs compete on detection velocity in a dual-use tool category. Every headline finding that Codex or Claude surfaces becomes adversarial training data. The same competitive pressure that will improve these scanners over the next 18 months also accelerates the sophistication of threat actors who are running identical API calls against identical open-source targets. Decision-makers who pilot these tools without locking down governance frameworks first are trading a known SAST blind spot for an unknown IP exposure. That is not an upgrade.
The Exploit
Opportunity
Security-forward enterprises that run structured pilots of both Claude Code Security and Codex Security before June 2026 will own the negotiating leverage that disappears at IPO. The research-preview window is a forcing function: teams that build internal benchmarks against their own production codebases now will enter procurement negotiations with vendor-specific performance data that neither OpenAI nor Anthropic can refute.
Risk
Security leaders who adopt either tool at scale without a governance framework first are the most exposed. The Checkmarx finding—75% false positive rate under real-world conditions—means a rushed rollout generates noise that desensitizes engineering teams, buries the two genuine critical findings in triage debt, and hands both labs proprietary intelligence about your codebase architecture before any contractual data protections exist.
The Move
The CISO owns this: commission a bounded 90-day pilot by June 2026 that runs both scanners against one production codebase with known vulnerability history, scoring recall and false positive rate against your internal ground truth. That benchmark becomes your contractual anchor when free tiers convert to paid—and your exit ramp if they don't perform.
First-Mover Advantage
Gains
Risks
Security leaders who adopt either tool at scale without a governance framework first are the most exposed. The Checkmarx finding—75% false positive rate under real-world conditions—means a rushed rollout generates noise that desensitizes engineering teams, buries the two genuine critical findings in triage debt, and hands both labs proprietary intelligence about your codebase architecture before any contractual data protections exist.
Window
Winners & Losers
Winners↑
Runtime protection and attack path analysis vendors
As LLM-based scanning commoditises static code analysis to zero, enterprise AppSec budgets are forecast to migrate toward runtime exploitability layers that reasoning scanners cannot replace. Vendors in this segment—offering runtime detection, attack path mapping, and exploitability context—inherit the procurement conversation that SAST vendors are losing. They should accelerate integrations with both Claude Code Security and Codex Security to position themselves as the essential triage and response layer sitting downstream of free scanning outputs.
AI governance and compliance platform providers
The Checkmarx validation finding—75% false positives in production conditions—and unresolved questions around training exclusion, data residency, and derived IP create an urgent governance gap that security leaders must close before piloting either scanner. Platforms that offer formal data-processing frameworks, segmented submission pipelines, and audit-grade controls are structurally positioned to capture budget as CISOs formalise governance around reasoning-based tools. The immediate play is to ship pre-built policy templates specifically scoped to Claude Code Security and Codex Security onboarding.
Remediation automation and patch orchestration vendors
Snyk's rebuttal identified the real bottleneck: finding vulnerabilities at LLM scale is now free, but fixing them across hundreds of repositories without breaking production is not. Vendors that automate remediation workflows, prioritise patches by runtime exploitability rather than raw CVSS, and integrate with CI/CD pipelines are positioned to absorb the AppSec budget displaced from SAST licensing. They should build direct connectors to both scanners' APIs now, before either lab ships native remediation features and closes the integration window.
Dual-use threat intelligence and adversarial research firms
Every headline CVE produced by Codex Security and Claude Code Security simultaneously becomes adversarial training data, compressing the window between discovery and exploitation across open-source dependencies that underpin enterprise applications. Threat intelligence firms that specialise in tracking LLM-assisted vulnerability discovery—and translating that into prioritised patch signals for enterprise clients—gain immediate relevance as CISOs shift triage from CVSS scores to real-world exploitability timelines. Publishing independent exploit-probability assessments for the 14 CVEs already assigned would establish category authority rapidly.
Losers↓
Legacy commercial SAST platform vendors
OpenAI and Anthropic have deliberately repriced static code analysis to zero for enterprise customers, and the structural argument for pattern-matching SAST—detecting known anti-patterns—has been publicly invalidated by two labs simultaneously finding vulnerability classes that traditional tools cannot reach. Vendors like Veracode and Checkmarx face a procurement conversation where security buyers can now point to free tools that outperform their paid tier on detection ceiling, even if real-world false positive rates remain a caveat. The defensive move is to reframe positioning around audit-grade reproducibility, pipeline governance, and compliance documentation—capabilities that probabilistic LLM scanners cannot currently deliver at enterprise certification standards.
Broad-scope enterprise AppSec suite vendors
Vendors selling integrated AppSec platforms that bundle SAST as a core value driver face margin compression across their highest-renewal line items, as the free research previews from two trillion-dollar labs erode the pricing power of the scanning component that anchors many multi-year contracts. The Cycode CTO's rebuttal—that IDE-embedded scanning does not constitute infrastructure—is technically accurate but strategically insufficient when enterprise security buyers are comparing a zero-cost reasoning scanner against a six-figure renewal. These vendors must urgently disaggregate their pricing architecture, move SAST to a loss-leader or free tier, and rebuild renewal justification entirely around orchestration, governance, and runtime behaviour layers that neither OpenAI nor Anthropic currently addresses.
Strategic Outlook
The SAST commoditization playbook has run before. Contrast Security gave away its community tier in 2019 to pressure Veracode and Checkmarx on price, and GitHub absorbed code scanning into the platform entirely when it acquired Semmle in 2019 and made CodeQL free for open-source in 2020. Both moves compressed margins across the category without eliminating it. What's different this time is the scale of the entities doing the pricing and their motivation for doing it now. OpenAI and Anthropic are both on IPO trajectories where enterprise security penetration becomes a key growth narrative for prospective public investors. The free scanner isn't a product decision—it's a pre-IPO land acquisition strategy, and the 18-month window before either company reaches public markets is exactly when lock-in gets built. This also lands inside a broader consolidation moment. CrowdStrike, Palo Alto, and Wiz have spent the last two years arguing that point solutions are dead and platform relationships are the only defensible position in enterprise security. Two AI labs entering with free, high-capability tools accelerates that argument by eliminating the standalone SAST license as a rational procurement choice—which conveniently redirects budget toward runtime, governance, and remediation layers where incumbents still hold pricing power.