OpenAI has officially launched GPT-5.5-Cyber, which is aimed at the security scene of the network, and has been incorporated into the Daybreak network defence plan. The model is primarily used to help security teams to identify and repair software gaps more quickly. In contrast, Mythos 5 of Anthropic is still on the lower line because of the emergency export restrictions previously imposed by the United States Government on national security grounds.

CyberGym up to 85.6%.

OpenAI announced a full release of GPT-5.5-Cyber on June 22. According to the CyberGym benchmark developed by the University of California at Berkeley, the model achieved 85.6 per cent of the controlled testing of 188 open-source projects and 1507 known software gaps.

On the same list, Mythos 5 of Anthropic scored 83.8 per cent, while Claude Opus 4.7 was 73.1 per cent open to a wider audience. In terms of the score gap, the two models did not stretch particularly far, but the market focus fell more on a completely different line of access for the two companies.

The Anthropic model still hasn't been restored.

Anthropic Mythos 5 and Fable 5 were offline on June 12th. Reports indicated that the Trump Government had issued emergency export control directives at that time on the grounds that the model involved a national security risk.

One of the issues raised by the Government was that the model could be circumvented by escapes. As Anthropic was unable to verify the citizenship of the user on a large scale and reliably, the company eventually chose to discontinue the two models globally instead of limiting access to only parts of the world.

Anthropic has previously emphasized on several occasions that Mythos is a highly capable and at the same time risky front-line model, and in the release material it is mentioned that, in the absence of sufficient restrictions, its network security capabilities may be misused. A few days later, this risk discussion quickly translated into practical regulatory action.

OpenAI to expand cooperation

As Anthropic continues to communicate with the United States Department of Commerce, OpenAI is expanding external cooperation with Daybreak. Public information indicates that the plan has established cybersecurity cooperation with relevant agencies in Australia, Canada, France, Germany, Japan, Korea and the EU.

  • 28 security companies participating in cooperative projects
  • Including CrowdStrike, Cisco, Cloudflare.
  • Capacity access is available for use by approved clients

According to OpenAI disclosure, its Codex Security tool has scanned more than 30 million code submissions since it went online in March, covering approximately 30,000 code libraries and recording more than half a million patches. The company simultaneously launched the “Patch the Planet” programme, the objective of which is to help repair the gaps in widely used open-source projects.

Still open to authentication users only

However, GPT-5.5-Cyber is not open to the public. According to OpenAI, the model is currently available only to certified security professionals and pre-deployment tests have been completed with United States federal agencies before going online.

It was mentioned that the agencies involved in the testing included the United States Centre for Artificial Intelligence Standards and Innovation and the Office of the Director of the National Network. Compared to Anthropic, the practice of OpenAI is not to relax restrictions, but to advance within control and coordinate with the Government in advance.

As of 23 June, Anthropic Fable 5 and Mythos 5 had not returned online and no clear recovery time had been given.