OpenAI Launches GPT-5.5-Cyber to Autonomously Patch Code and Outperform Anthropic

OpenAI has announced a major expansion of its Daybreak cybersecurity initiative, showcasing an updated Codex Security plugin, the launch of its full GPT-5.5-Cyber model, and an extensive network of over twenty-five enterprise security firms and several global governments[1]. In a notable escalation of the ongoing competitive rivalry between the world's leading artificial intelligence laboratories, OpenAI reports that its new GPT-5.5-Cyber model has successfully outperformed Anthropic's highly acclaimed Mythos model on a critical industry cybersecurity benchmark[2]. The announcement marks a fundamental shift in how tech companies utilize artificial intelligence for defense, moving the strategic objective away from merely identifying software vulnerabilities and toward automatically generating, testing, and merging functional security patches[1][3]. By accelerating the end-to-end remediation loop at machine speed, this updated suite of tools aims to resolve the systemic backlogs of security flaws that threaten modern enterprise and open-source software ecosystems alike[4].

The core of OpenAI's technical claim rests on the newly released GPT-5.5-Cyber model's performance on CyberGym, a leading public benchmark designed to evaluate the capability of autonomous AI agents on complex, multi-step security tasks[2][5]. In recent evaluations, the full version of GPT-5.5-Cyber achieved a state-of-the-art score of 85.6 percent, surpassing the 81.8 percent achieved by the standard GPT-5.5 model, as well as the 83.1 percent peak score previously held by Anthropic's Claude Mythos[2][5]. Beyond CyberGym, the specialized security model displayed significant generational improvements across other evaluations, posting a score of 39.5 percent on ExploitGym compared to the standard model's 25.95 percent, and reaching 69.8 percent on the advanced SEC-bench Pro evaluation[6]. Independent testing by government-backed research bodies has corroborated these rapid advancements[7]. The United Kingdom's AI Safety Institute recently revealed that early checkpoints of both GPT-5.5 and Claude Mythos were the first AI systems capable of autonomously executing "The Last Ones," a highly complex, 32-step corporate network attack simulation that simulates reconnaissance, lateral movement, and privilege escalation, which typically requires twenty hours of manual effort from an expert human engineer[8][9].

These soaring benchmark figures reflect a broader, more critical evolution in the field of digital defense[4]. For years, the primary challenge for security organizations was finding hidden flaws within code, but the introduction of frontier AI models has inverted this dynamic[3][2]. AI has made finding vulnerabilities so efficient that software maintainers and application security teams are now inundated with a massive volume of automated bug reports[3][10]. Recognizing that a sheer volume of alerts does not make systems safer, OpenAI designed the updated Codex Security plugin to close the loop from discovery to patch generation[1][4]. The plugin effectively functions as an automated security engineer integrated directly into developer workflows[1]. Since its initial research preview, the Codex Security plugin has scanned over thirty million commits across more than thirty thousand codebases, with over five hundred thousand findings automatically resolved and an additional seventy thousand fixes manually confirmed by human developers[1][6]. By allowing developers to analyze code alongside threat models, validate findings in isolated environments, and generate codebase-specific patches for immediate review, the updated plugin reduces critical hours of analysis to mere minutes[4][3].

To demonstrate the real-world viability of this automated patching pipeline, OpenAI has introduced a new open-source security program called Patch the Planet[6][10]. Launched in partnership with cybersecurity research firm Trail of Bits, HackerOne, and Calif, the initiative couples OpenAI's frontier models with human oversight to directly secure the foundational infrastructure of the modern internet[6][10][11]. During its first week of operation, Trail of Bits deployed its entire security research organization to collaborate with maintainers across nineteen critical open-source projects, including Python, Go, cURL, Sigstore, NATS Server, pyca/cryptography, and aiohttp[3][10][12]. Utilizing GPT-5.5-Cyber and the Codex Security plugin, engineers scanned security-relevant components across millions of lines of code, identifying hundreds of legitimate bugs and successfully merging dozens of functional patches[6][12]. Additionally, the model demonstrated its deep reasoning capabilities by generating verified proof-of-concepts for eight kernel pointer information leaks and twenty-four local privilege escalation vulnerabilities within the Linux kernel[6]. This collaborative approach ensures that open-source maintainers, who are frequently underfunded volunteers, are not overwhelmed by raw vulnerability reports but are instead presented with vetted, tested, and ready-to-implement code fixes[10][11].

Deploying such powerful capabilities requires a carefully calibrated safety and distribution model[4]. Because the same intelligence required to generate defensive patches can be subverted for offensive cyber operations, OpenAI is strictly limiting access to the full GPT-5.5-Cyber model[4][13]. The technology is being distributed through a restricted "Trusted Access for Cyber" program designed specifically for verified defenders, national security agencies, and selected enterprise partners[4][6]. This safety-first approach aligns closely with Anthropic's decision to gate its Claude Mythos model through its highly restricted Project Glasswing in collaboration with the United States government[14][15]. To build a robust protective perimeter around these tools, OpenAI has also launched the Daybreak Cyber Partner Program[2][6]. This ecosystem network incorporates over twenty-five prominent global security firms, including industry giants such as Accenture, Akamai, Check Point, Cisco, Cloudflare, CrowdStrike, IBM, and Palo Alto Networks[6]. By integrating these advanced defensive models into the existing security systems and workflows that enterprises already trust, the initiative aims to scale the protective benefits of frontier AI without exposing raw dual-use capabilities to malicious actors[4].

The launch of GPT-5.5-Cyber and the expansion of the Daybreak initiative carry profound implications for the global AI and cybersecurity industries[16]. As offensive AI capabilities continue to expand at an unprecedented pace, the traditional window of opportunity for defenders to respond to zero-day threats is shrinking rapidly[16][17]. The transition toward autonomous, AI-driven remediation is no longer a futuristic aspiration but a critical necessity[4][2]. By outperforming established benchmarks and proving its utility through active, merged code contributions in major open-source repositories, OpenAI's latest model signals that the balance of power in digital security may finally be shifting back to defenders[4][6]. In an era where software supply chains are increasingly complex and heavily targeted by sophisticated adversaries, the successful deployment of guarded, highly capable defensive AI agents could establish a new standard for national security and global digital resilience[4][18].