OpenAI has unveiled Codex Security, a new application security agent designed to streamline the discovery and remediation of vulnerabilities. Previously named Aardvark, this tool is now available in a research preview and aims to address the inefficiencies of manual security reviews. By leveraging advanced AI models integrated with automated validation processes, Codex Security allows development teams to deploy secure code more swiftly while significantly reducing the noise associated with triage.
Traditional AI security solutions often inundate security teams with low-impact alerts and false positives, creating more confusion than clarity. Codex Security counters this issue by thoroughly analyzing a code repository to grasp its specific structure, which enables the generation of a tailored threat model. This model elucidates the system’s functions, trust relationships, and points of vulnerability, allowing the agent to identify and prioritize vulnerabilities based on their likely real-world impact.
A key feature of Codex Security is its capability to pressure-test findings in sandboxed environments, yielding high-confidence reporting. This includes the ability to generate working proof-of-concept exploits, which enhances the tool’s reliability. In a recent beta phase, the system demonstrated remarkable efficacy, achieving an 84 percent reduction in overall noise, a 90 percent decline in over-reported severity findings, and a 50 percent decrease in false-positive rates. During a 30-day assessment, Codex Security scanned over 1.2 million commits across various external repositories, identifying 792 critical and 10,561 high-severity findings while maintaining minimal noise. Notably, critical issues were found in less than 0.1 percent of the scanned commits, underscoring the tool’s ability to manage vast volumes of code efficiently.
Early adopters of Codex Security, including NETGEAR, reported seamless integration into their development environments. Chandan Nandakumaraiah, Head of Product Security at NETGEAR, noted that the tool’s comprehensive findings were akin to having a seasoned product security researcher augmenting their team.
Codex Security’s core functionalities include threat modeling, which aligns security checks with actual system exposure; issue validation, which tests vulnerabilities to minimize false positives; and automated patching, which proposes tailored fixes aimed at preventing software regressions. The system also features adaptive learning, which refines its threat model based on team feedback regarding the criticality of findings, further reducing the triage burden and enhancing precision.
In a bid to bolster open-source software supply chain security, OpenAI is deploying Codex Security to assist open-source maintainers who often contend with an influx of low-quality bug reports. By prioritizing actionable, high-confidence vulnerabilities, Codex Security has already identified significant flaws in notable open-source projects. These include a critical security vulnerability in the portable version of OpenSSH, a high-severity issue in GnuTLS, and a repository exposure problem in GOGS. Notably, vulnerabilities in Thorium have been tracked under CVE-2025-35430.
Thus far, the tool has been instrumental in uncovering 14 CVEs across various projects, including PHP, libssh, and Chromium. To further support the developer community, OpenAI has introduced “Codex for OSS,” which provides free ChatGPT Pro accounts, code review tools, and access to Codex Security for open-source maintainers. Codex Security is accessible in research preview via the Codex web interface, offering free usage for the initial month.
As OpenAI continues to refine Codex Security, its potential impact on both proprietary and open-source software ecosystems could reshape best practices in vulnerability management, enabling more secure and efficient development cycles. The tool’s advanced capabilities not only enhance security workflows but also hold promise for fostering a more resilient software landscape.
See also
OpenAI Launches Codex Security, AI Agent That Identifies and Fixes Code Vulnerabilities
IBM Report Reveals 44% Surge in AI-Driven Cyberattacks Targeting Vulnerable Systems
24.3% of Companies Paid Ransoms in 2025 as AI-Driven Cyberattacks Escalate
Iran Conflict Sparks Surge in AI-Driven Cyberattacks, Heightening Global Risks for Insurers
AI-Driven Cyberattacks Surge as 90% Target Vulnerable Third-Party Software
