GPT-5.5-Cyber Launches with Record 85.6 Percent CyberGym Score Under Daybreak Initiative

[VoidSentinel]

OpenAI launched the full version of GPT-5.5-Cyber on June 22nd as the centrepiece of its expanded Daybreak cybersecurity initiative and the benchmark numbers are legitimately impressive. The model scored 85.6 percent on CyberGym, the highest single-model result ever recorded on that benchmark, compared to 81.8 percent for standard GPT-5.5. The model also scored 39.5 percent on ExploitGym, against 25.95 percent for the standard model. These are substantial improvements that demonstrate real capability uplift specifically tuned for offensive and defensive security tasks.

The model is not a public API product. Access is tightly gated through OpenAI's Trusted Access for Cyber program which currently includes Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks and Zscaler. Government agencies, enterprise security teams and academic researchers can also apply. This is clearly designed to prevent the kind of scenario that triggered the Anthropic export ban, where broad deployment of a highly capable cybersecurity model attracted government alarm.

The irony noted widely in the security community is that Anthropic's Mythos 5 scored 83.8 percent on the same benchmark yet was banned, while GPT-5.5-Cyber at 85.6 percent is being distributed to private companies under a controlled access programme. The approaches are genuinely different in philosophy. OpenAI's restricted access model gives them more control over who uses the capability and for what purposes, while Anthropic's broad deployment aimed to make defenders everywhere more capable.

Daybreak Securing The Worldopenai.com

[TheLegendJohn32]

85.6 percent on CyberGym is remarkable but what I want to know is what the 14.4 percent failure looks like. In offensive security the misses can be as dangerous as the hits

[HeartbreakKid]

The list of Trusted Access companies reads like a who's who of corporate cybersecurity. No academic institutions on the public list which is odd given the stated research access pathway

[Blue Sasha]

The comparison to Mythos 5 at 83.8 percent is deeply uncomfortable for the government's logic in banning Anthropic but not OpenAI. Capability alone cannot be the criterion

[EdgeRatedR]

Restricted deployment to vetted partners is genuinely different from open API access and I think that distinction matters for national security arguments. Anthropic chose breadth and got penalised for it

[DataStream Luca]

ExploitGym at 39.5 versus 25.95 for standard GPT-5.5 is the scary number. That is a 52 percent relative improvement in exploit generation capability. Red team firms need this but so do bad actors

[WhatUQuant]

Codex Security scanning 30 million commits across 30,000 codebases and having 70,000 findings fixed is a much better story for public safety than these raw benchmark numbers

[Finley_19]

I have submitted an application to the Trusted Access program. The process is not described publicly anywhere which is its own transparency problem

[LurkingLegend]

The real question is whether this model will eventually become available more broadly once the regulatory framework around AI cybersecurity capabilities matures. The benchmarks suggest it is very capable