OpenAI’s New GPT5.5-Cyber Beats Anthropic’s Mythos, But No Ban Yet

OpenAI has unveiled the full version of its cybersecurity-focused AI model GPT-5.5-Cyber. The model is part of the company’s own cyber-defense program “Daybreak” and is explicitly aimed not at the general public but at verified security professionals. While Anthropic’s competing flagship models Mythos 5 and Fable 5 are currently offline by order of the US government, OpenAI’s new model continues to operate under access controls coordinated with the government.

That GPT-5.5 may be more capable – and therefore, from a cybersecurity standpoint, also more dangerous – than Claude Mythos is something experts have already pointed out (Trending Topics reported).

What the model can do

GPT-5.5-Cyber is meant to help security teams find and fix software vulnerabilities faster. According to OpenAI, the bottleneck in cyber defense is currently shifting: the main problem is no longer finding vulnerabilities, but actually remediating them. The model is designed to analyze large codebases, trace whether vulnerable code is even reachable, validate vulnerabilities in controlled environments, and develop and test patches – with humans as the final point of control.

On CyberGym, a benchmark developed at UC Berkeley that tests whether AI agents can reproduce known vulnerabilities in software environments, the new model reaches 85.6 percent according to OpenAI. For comparison, the company cites 81.8 percent for the standard model GPT-5.5 – by its own account the highest CyberGym score OpenAI has measured to date for a single model. On two further benchmarks, the company reports gains: 39.5 percent (versus 25.95 percent for GPT-5.5) on ExploitGym, which tests whether agents can turn known vulnerabilities into working exploits, and 69.8 percent (versus 63.1 percent) on SEC-bench Pro.

The comparison with Anthropic

The comparison with the direct competitor is the explosive part. On the CyberGym leaderboard, OpenAI’s new model leads with 85.6 percent, ahead of Anthropic’s Mythos 5, which comes in at 83.8 percent. Anthropic’s more widely available model, Claude Opus 4.7, reaches 73.1 percent.

A gap of less than two percentage points would be hardly remarkable on its own. The context, however, is: Mythos 5 and Fable 5 were shut down on 12 June following an emergency export-control directive from the Trump administration, justified on national-security grounds. The trigger, according to reports, was a jailbreak – a technique for bypassing a model’s built-in safety restrictions. Because Anthropic had no reliable way to verify the nationality of its users at scale, the company disabled both models worldwide for everyone. As of 23 June, Fable 5 and Mythos 5 remained offline – with no official restoration date from Anthropic or the US Department of Commerce.

OpenAI is taking a similarly restrictive access approach with its launch, but coordinated it with the US government in advance. GPT-5.5-Cyber is available only to vetted security professionals; before launch, OpenAI says it ran tests with federal agencies, including the Center for AI Standards and Innovation (CAISI) and the Office of the National Cyber Director (ONCD). For most defenders, however, GPT-5.5 combined with “Trusted Access for Cyber” and the Codex Security tool is the appropriate starting point; the Cyber variant is aimed at those whose authorized work requires the most advanced capabilities and a “more permissive” model behavior.

Partnerships and ecosystem

Alongside the model, OpenAI is expanding its cyber ecosystem. The Daybreak Cyber Partner Program is meant to let security providers integrate GPT-5.5 into their own products and services, while direct model access stays with the partners. The named partners include CrowdStrike, Cisco, Cloudflare, Palo Alto Networks, IBM, Akamai, Check Point, Fortinet, SentinelOne and Zscaler, among others.

At the governmental level, OpenAI says it has signed “Trusted Access for Cyber” partnerships over the past month with Australia, Canada, France, Germany, Japan, the Republic of Korea, and EU institutions such as the EU cybersecurity agency ENISA; a collaboration with the UK government also exists. Added to this is the “Patch the Planet” initiative, launched together with Trail of Bits, HackerOne and other actors to close security gaps in widely used open-source software – according to the company, more than 30 projects have taken part, including cURL, Go, Python and pyca/cryptography. The Codex Security tool, the company says, has scanned more than 30 million commits across over 30,000 codebases since launching in March and logged more than 500,000 fixed vulnerabilities.

Background: the earlier Pentagon dispute

The current contrast between the two companies builds on an earlier confrontation that should be separated from the June export ban both in time and in substance. As early as late February and early March 2026, negotiations between Anthropic and the US Department of Defense over the use of Anthropic’s models in classified environments had collapsed. Anthropic insisted on red lines – no use for autonomous weapons systems and no domestic mass surveillance. In response, President Trump directed federal agencies to phase out Anthropic’s technology over a six-month transition period; Secretary of Defense Pete Hegseth designated the company a “supply-chain risk.” Anthropic called that designation “unlawful and politically motivated.”

A few hours later, OpenAI announced a contract of its own for deploying its models in classified environments. CEO Sam Altman conceded that the deal had been “definitely rushed,” but stressed that OpenAI had not simply accepted the terms Anthropic had rejected. According to OpenAI’s account, Anthropic had named specific contractual prohibitions as a condition, whereas OpenAI relied on existing laws and policies – an approach critics rated as legally softer. Observers such as MIT Technology Review read this as Anthropic pursuing a more morally strict course and OpenAI a more pragmatic one. Anthropic, which previously held a 200-million-dollar contract with the Department of Defense and whose Claude models were already embedded in certain classified systems, later returned to the negotiating table.

The export controls imposed in June against Mythos 5 and Fable 5 are a separate matter. Anthropic says it continues to negotiate with the Department of Commerce while simultaneously suing the Trump administration. OpenAI, for its part, points to an “ongoing dialogue” with the US government about its cyber approach and upcoming model releases.