In short
- OpenAI released GPT-5.5-Cyber, a version designed to help find and fix software problems faster than previous versions.
- It surpasses the Anthropic Mythos on major landmarks, while the Mythos faces a US government ban due to national security concerns.
- The model is being released to trusted defenders with controls, unlike the restricted Anthropic system.
OpenAI’s cybersecurity model just beat the US government’s Anthropic Mythos AI version offline — and it’s still working.
On June 22, OpenAI announced complete installation of GPT-5.5-Cyber as part of its Daybreak cyber Defense program. On CyberGym – a benchmark developed at UC Berkeley that provides AI agents with a risk assessment of 1,507 known programs from 188 open source projects and calculates how many they can increase in a controlled environment – the modified version reached 85.6%.
Anthropic’s Mythos 5 sits at 83.8% on the same scoreboard. Claude Opus 4.7, the most widely available version of Anthropic, scored 73.1%.
A difference of two points on the benchmark would not be surprising. The content here is not. Anthropic’s Mythos 5 and Fable 5 were was released on June 12 After Donald Trump’s administration issued an emergency export ban on national security grounds.
The government described it as a jailbreak – a way to bypass the security boundaries of an AI system, similar to finding a technical key that opens a high-security door. Anthropic didn’t have a reliable way to verify the user version on a large scale, so it disabled both versions for everyone, everywhere.
Some of the damage was self-inflicted. Anthropic spent months describing Mythos as one of the most capable and dangerous types of AI ever built, warning in its launch statement that its cyber security capabilities could cause serious problems without proper controls. Anthropic CEO Dario Amodei published story on June 10 to compare the AI models of borders and planes that security forces must deploy in the event of a failed search.
A few days later, the government suspended Anthropic flights.
It wasn’t the only alarm that week. Anthropic was it’s already burnt on a hidden filter in Fable 5 that silently degraded the model’s output to users it suspected of creating an AI competitor – without telling them – and was forced to apologize and change the policy.
Another playbook
As Anthropic negotiates with the Commerce Department and continues its lawsuit against the Trump administration, OpenAI is expanding its reach. Daybreak has signed cybersecurity agreements with Australia, Canada, France, Germany, Japan, South Korea, and the EU. including European Union Agency for Cybersecurity.
Twenty-eight security companies – including CrowdStrike, Cisco, and Cloudflare – have joined its Cyber Partner program to implement GPT-5.5 in their products for rated customers. On the OpenAI blog, the Codex Security tool has analyzed more than 30 million objects across 30,000 codebases and identified more than 500,000 threats since its launch in March.
The company is also expanding its partner program so security businesses can integrate these into their tools, and launched “Patch the Planet,” a way to help fix vulnerabilities in widely used open source projects.
That said, GPT-5.5-Cyber is not for general use. It’s only available to certified security professionals, and OpenAI underwent pre-deployment testing by government agencies — including the Center for AI Standards and Innovation and the Office of the National Cyber Director — before launch. This is a similar approach that Anthropic tried with Mythos, but OpenAI cleared its way with the government first.
As of June 23, Fable 5 and Mythos 5 remain offline—an eleven-day suspension with no recovery date from Anthropic or the Commerce Department.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
