OpenAI has launched Codex Security, an AI agent built on its Codex coding system to help developers find and fix security flaws. Released on March 6, 2026, it is available as a research preview to ChatGPT Pro, Enterprise, Business, and Education customers through the Codex web interface, with the first month free.
What Codex Security does
Codex Security analyzes a code repository and produces a detailed, natural-language description of how the application works, where it is strongest, and where security vulnerabilities may exist. Each potential flaw is tested in a sandbox to filter out false positives, and the findings are ranked by likely severity and real-world impact. The agent then generates possible fixes for each issue, including the relevant code and a plain-language explanation, and developers can approve a patch directly from the interface and push it to production. According to OpenAI, most AI security tools surface mainly low-impact findings and false positives, which forces security teams to spend heavily on triage; combining agentic reasoning with automated validation, it says, is intended to deliver higher-confidence findings and actionable fixes. A related discussion of disclosing security flaws responsibly appears in this piece on responsible vulnerability disclosure.
Early results
Over its first 30 days, OpenAI says Codex Security scanned more than 1.2 million commits across external repositories and identified 792 critical and 10,561 high-severity issues. Among the open-source projects flagged were widely used components, and 14 of the findings were serious enough to be logged as entries in the CVE (Common Vulnerabilities and Exposures) database maintained by the MITRE Corporation. The company also reports that, since the earlier private test, it has reduced noise and false positives by more than 50 percent.
Background and competition
Codex Security was first offered last year as a private beta called Aardvark, tested with a small group of customers. OpenAI says it has begun enlisting an initial group of open-source maintainers to try the tool, with plans to expand the program. The launch comes shortly after Anthropic added competing security-scanning capabilities to its Claude Code tool, which similarly scans codebases, identifies vulnerabilities, and suggests fixes — a sign of growing competition in AI-assisted application security.
What to watch
Automated security tooling is promising but not a complete solution. Sandbox validation reduces false positives without eliminating them, and severity rankings still benefit from human review before a patch is shipped. Because the tool generates code fixes, those patches should be examined and tested rather than approved automatically, particularly for critical systems. As a research preview, the feature is also likely to change, so its real value will become clearer as more maintainers and security teams put it to use on production codebases.
