OpenAI launched Codex Security on March 6, entering the application security market that Anthropic had disrupted 14 days earlier with Claude Code Security. Both scanners use LLM reasoning instead of pattern matching. Both proved that traditional static application security testing (SAST) tools are structurally blind to entire vulnerability classes. The enterprise security stack is caught in the middle.

Anthropic and OpenAI independently released reasoning-based vulnerability scanners, and both found bug classes that pattern-matching SAST was never designed to detect. The competitive pressure between two labs with a combined private-market valuation exceeding $1.1 trillion means detection quality will improve faster than any single vendor can deliver alone.

Neither Claude Code Security nor Codex Security replaces your existing stack. Both tools change procurement math permanently. Right now, both are free to enterprise customers. The head-to-head comparison and seven actions below are what you need before the board of directors asks which scanner you are piloting and why.

How Anthropic and OpenAI reached the same conclusion from different architectures

Anthropic published its zero-day research on February 5 alongside the release of Claude Opus 4.6. Anthropic said Claude Opus 4.6 found more than 500 previously unknown high-severity vulnerabilities in production open-source codebases that had survived decades of expert review and millions of hours of fuzzing.

In the CGIF library, Claude discovered a heap buffer overflow by reasoning about the LZW compression algorithm, a flaw that coverage-guided fuzzing could not catch even with 100% code coverage. Anthropic shipped Claude Code Security as a limited research preview on February 20, available to Enterprise and Team customers, with free expedited access for open-source maintainers. Gabby Curtis, Anthropic’s communications lead, told VentureBeat in an exclusive interview that Anthropic built Claude Code Security to make defensive capabilities more widely available.

OpenAI’s numbers come from a different architecture and a wider scanning surface. Codex Security evolved from Aardvark, an internal tool powered by GPT-5 that entered private beta in 2025. During the Codex Security beta period, OpenAI’s agent scanned more than 1.2 million commits across external repositories, surfacing what OpenAI said were 792 critical findings and 10,561 high-severity findings. OpenAI reported vulnerabilities in OpenSSH, GnuTLS, GOGS, Thorium, libssh, PHP, and Chromium, resulting in 14 assigned CVEs. Codex Security’s false positive rates fell more than 50% across all repositories during beta, according to OpenAI. Over-reported severity dropped more than 90%.

Checkmarx Zero researchers demonstrated that moderately complicated vulnerabilities sometimes escaped Claude Code Security’s detection. Developers could trick the agent into ignoring vulnerable code. In a full production-grade codebase scan, Checkmarx Zero found that Claude identified eight vulnerabilities, but only two were true positives. If moderately complex obfuscation defeats the scanner, the detection ceiling is lower than the headline numbers suggest. Neither Anthropic nor OpenAI has submitted detection claims to an independent third-party audit. Security leaders should treat the reported numbers as indicative, not audited.

Merritt Baer, CSO at Enkrypt AI and former Deputy CISO at AWS, told VentureBeat that the competitive scanner race compresses the window for everyone. Baer advised security teams to prioritize patches based on exploitability in their runtime context rather than CVSS scores alone, shorten the window between discovery, triage, and patch, and maintain software bill of materials visibility so they know instantly where a vulnerable component runs.

Different methods, almost no overlap in the codebases they scanned, yet the same conclusion. Pattern-matching SAST has a ceiling, and LLM reasoning extends detection past it. When two competing labs distribute that capability at the same time, the dual-use math gets uncomfortable. Any financial institution or fintech running a commercial codebase should assume that if Claude Code Security and Codex Security can find these bugs, adversaries with API access can find them, too.

Baer put it bluntly: open-source vulnerabilities surfaced by reasoning models should be treated closer to zero-day class discoveries, not backlog items. The window between discovery and exploitation just compressed, and most vulnerability management programs are still triaging on CVSS alone.

What the vendor responses prove

Snyk, the developer security platform used by engineering teams to find and fix vulnerabilities in code and open-source dependencies, acknowledged the technical breakthrough but argued that finding vulnerabilities has never been the hard part. Fixing them at scale, across hundreds of repositories, without breaking anything. That is the bottleneck. Snyk pointed to research showing AI-generated code is 2.74 times more likely to introduce security vulnerabilities compared to human-written code, according to Veracode’s 2025 GenAI Code Security Report. The same models finding hundreds of zero-days also introduce new vulnerability classes when they write code.

Cycode CTO Ronen Slavin wrote that Claude Code Security represents a genuine technical advancement in static analysis, but that AI models are probabilistic by nature. Slavin argued that security teams need consistent, reproducible, audit-grade results, and that a scanning capability embedded in an IDE is useful but does not constitute infrastructure. Slavin’s position: SAST is one discipline within a much broader scope, and free scanning does not displace platforms that handle governance, pipeline integrity, and runtime behavior at enterprise scale.

“If code reasoning scanners from major AI labs are effectively free to enterprise customers, then static code scanning commoditizes overnight,” Baer told VentureBeat. Over the next 12 months, Baer expects the budget to move toward three areas.

1. Runtime and exploitability layers, including runtime protection and attack path analysis.

2. AI governance and model security, including guardrails, prompt injection defenses, and agent oversight.

3. Remediation automation. “The net effect is that AppSec spending probably doesn’t shrink, but the center of gravity shifts away from traditional SAST licenses and toward tooling that shortens remediation cycles,” Baer said.

Seven things to do before your next board meeting

1. Run both scanners against a representative codebase subset. Compare Claude Code Security and Codex Security findings against your existing SAST output. Start with a single representative repository, not your entire codebase. Both tools are in research preview with access constraints that make full-estate scanning premature. The delta is your blind spot inventory.

2. Build the governance framework before the pilot, not after. Baer told VentureBeat to treat either tool like a new data processor for the crown jewels, which is your source code. Baer’s governance model includes a formal data-processing agreement with clear statements on training exclusion, data retention, and subprocessor use, a segmented submission pipeline so only the repos you intend to scan are transmitted, and an internal classification policy that distinguishes code that can leave your boundary from code that cannot. In interviews with more than 40 CISOs, VentureBeat found that formal governance frameworks for reasoning-based scanning tools barely exist yet. Baer flagged derived IP as the blind spot most teams have not addressed. Can model providers retain embeddings or reasoning traces, and are those artifacts considered your intellectual property? The other gap is data residency for code, which historically was not regulated like customer data but increasingly falls under export control and national security review.

3. Map what neither tool covers. Software composition analysis. Container scanning. Infrastructure-as-code. DAST. Runtime detection and response. Claude Code Security and Codex Security operate at the code-reasoning layer. Your existing stack handles everything else. That stack’s pricing power is what shifted.

4. Quantify the dual-use exposure. Every zero-day Anthropic and OpenAI surfaced lives in an open-source project that enterprise applications depend on. Both labs are disclosing and patching responsibly, but the window between their discovery and your adoption of those patches is exactly where attackers operate. AI security startup AISLE independently discovered all 12 zero-day vulnerabilities in OpenSSL’s January 2026 security patch, including a stack buffer overflow (CVE-2025-15467) that is potentially remotely exploitable without valid key material. Fuzzers ran against OpenSSL for years and missed every one. Assume adversaries are running the same models against the same codebases.

5. Prepare the board comparison before they ask. Claude Code Security reasons about code contextually, traces data flows, and uses multi-stage self-verification. Codex Security builds a project-specific threat model before scanning and validates findings in sandboxed environments. Each tool is in research preview and requires human approval before any patch is applied. The board needs side-by-side analysis, not a single-vendor pitch. When the conversation turns to why your existing suite missed what Anthropic found, Baer offered framing that works at the board level. Pattern-matching SAST solved a different generation of problems, Baer told VentureBeat. It was designed to detect known anti-patterns. That capability still matters and still reduces risk. But reasoning models can evaluate multi-file logic, state transitions, and developer intent, which is where many modern bugs live. Baer’s board-ready summary: “We bought the right tools for the threats of the last decade; the technology just advanced.”

6. Track the competitive cycle. Both companies are heading toward IPOs, and enterprise security wins drive the growth narrative. When one scanner misses a blind spot, it lands on the other lab’s feature roadmap within weeks. Both labs ship model updates on monthly cycles. That cadence will outrun any single vendor’s release calendar. Baer said that running both is the right move: “Different models reason differently, and the delta between them can reveal bugs neither tool alone would consistently catch. In the short term, using both isn’t redundancy. It’s defense through diversity of reasoning systems.”

7. Set a 30-day pilot window. Before February 20, this test did not exist. Run Claude Code Security and Codex Security against the same codebase and let the delta drive the procurement conversation with empirical data instead of vendor marketing. Thirty days gives you that data.

Fourteen days separated Anthropic and OpenAI. The gap between the next releases will be shorter. Attackers are watching the same calendar.

Security | VentureBeat – ​Read More