# OpenAI Releases GPT-5.5-Cyber Cybersecurity Model in Limited Preview

> The specialized variant offers permissive capabilities for verified security teams while incorporating stricter verification to prevent misuse.

*Published 2026-06-23 · By Marcus Vance*

GPT-5.5-Cyber is a cybersecurity-specialized variant of OpenAI's GPT-5.5 model designed for permissive use in authorized defensive workflows.

OpenAI introduced GPT-5.5-Cyber on May 7, 2026, through its Trusted Access for Cyber program. The release targets defenders responsible for critical infrastructure. This limited preview grants access to specialized workflows that require greater model permissiveness than standard offerings.

## What background information explains the need for a cyber-specific model?

Frontier models have shown increasing capability in cybersecurity tasks. Previous versions demonstrated strong performance but required additional safeguards. OpenAI developed GPT-5.5-Cyber to balance capability with controlled access. The approach builds on earlier evaluations that highlighted the potential for both defensive and offensive applications.

Anthropic has pursued a different path with its Mythos series. External limits on Mythos 5 reflect a more restrictive stance on cyber capabilities. In contrast, OpenAI opted for a partner program that vets users before granting access to the more permissive variant.

## How does GPT-5.5-Cyber perform compared to competing models?

The model recorded 85.6 percent on the CyberGym benchmark according to reporting from Axios. This result exceeds the 81.8 percent achieved by the base GPT-5.5 and the 83.8 percent from Anthropic's Mythos 5. An updated version further improves on vulnerability reproduction tasks.

Benchmark Performance ComparisonModelCyberGym ScoreExpert Tasks Pass RateKey FeatureGPT-5.5-Cyber85.6%N/APermissive for authorized useGPT-5.581.8%71.4%Base modelMythos 583.8%68.6%External limits applied

On expert-level cyber tasks, the base GPT-5.5 model achieved an average pass rate of 71.4 percent per AISI evaluations. This placed it among the strongest models tested. AISI noted that GPT-5.5 solved one multi-step cyber-attack simulation end-to-end, a rare accomplishment.

## What technical specifics define the access controls for GPT-5.5-Cyber?

Access requires participation in the Daybreak Cyber Partner Program. Verified security vendors and defenders receive the model through this channel. Stronger verification processes and account-level controls accompany the release. These measures aim to restrict use to authorized red teaming, penetration testing, and controlled validation.

- Apply for verification through the Trusted Access for Cyber program
- Receive approval for specialized workflows
- Integrate the model with partner security platforms
- Monitor usage with account-level controls
- Report on authorized defensive activities

The list above outlines the structured process for gaining access. Partners such as Cisco and CrowdStrike provide the integration layer that embeds the model into existing defensive tools. This setup allows faster operations without compromising enterprise trust frameworks.

> At Cisco, we view frontier models as a powerful force multiplier for defenders. Models like GPT-5.5 are fundamentally changing the velocity of our operations, enabling us to move faster on everything from incident investigation to proactive exposure reduction. But speed cannot be traded for trust. The true value of this technology isn't found in the model alone, but in the enterprise-ready framework we wrap around it.Anthony Grieco, SVP, Chief Security & Trust Officer, Cisco

## What market and stakeholder implications arise from this release?

The partnerships with five major security vendors signal broader industry adoption of frontier models in defensive roles. Defenders gain tools that accelerate incident response and exposure reduction. However, the limited preview nature means only a subset of organizations will access the capabilities initially.

Stakeholders in critical infrastructure sectors stand to benefit most. The model supports proactive measures that traditional tools may not match in speed. Yet the requirement for verification creates a barrier that ensures responsible deployment.

## How have experts and organizations reacted to the GPT-5.5-Cyber announcement?

Cisco's leadership emphasized the combination of model power and enterprise controls. The statement highlights that trust remains paramount even as capabilities advance. Other partners including Cloudflare and Palo Alto Networks have similarly positioned the technology as a multiplier for their platforms.

AISI evaluations underscore the technical strength while noting the need for continued monitoring of dual-use potential. The agency identified GPT-5.5 as one of the strongest models tested on cyber tasks.

## What developments are expected next in frontier cyber models?

OpenAI indicated plans to expand the Trusted Access for Cyber program. Additional updates to GPT-5.5-Cyber may further enhance benchmark performance. The company continues to refine verification mechanisms to scale access responsibly.

Competitors may respond with their own specialized variants or adjusted policies. The field is likely to see increased focus on controlled release mechanisms for high-risk capability areas like cybersecurity.

## Sources

1. [GPT-5.5-Cyber achieved 85.6% on CyberGym benchmark outperforming GPT-5.5 at 81.8% and Anthropic Mythos 5 at 83.8%](https://www.axios.com/2026/06/22/openai-rolls-out-more-capable-version-of-cyber-model)
2. [GPT-5.5 achieves average pass rate of 71.4% on Expert-level cyber tasks and is the second model to solve one multi-step cyber-attack simulation end-to-end compared to 68.6% for Mythos Preview](https://www.aisi.gov.uk/blog/our-evaluation-of-openais-gpt-5-5-cyber-capabilities)
3. [Today, we are rolling out GPT-5.5-Cyber in limited preview to defenders responsible for securing critical infrastructure to support specialized cybersecurity workflows](https://openai.com/index/gpt-5-5-with-trusted-access-for-cyber/)

---
Source: https://aiintelreport.com/frontier-models/openai-gpt-5-5-cyber-cybersecurity
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt