GPT-5.5-Cyber¶

Development line: project:gpt-5-5-cyber · thread gpt-5-5-cyber
Last event: 2026-06-23 · 1 dated since 2026-06-23 · Researched: 2026-09-05 · confidence: high

What it is¶

GPT-5.5-Cyber is a legacy OpenAI model for approved defenders doing authorized red teaming, exploit validation, penetration testing, and controlled vulnerability research.

Capabilities:

• Cyber workflows: runs specialized authorized security operations.
• Refusals: lowered for higher-risk defensive tasks.
• API identifier: uses gpt-5.5-cyber-preview in the Responses API.

Prior GPT-5.5-Cyber approval does not grant Daybreak Red. New advanced access requires separate approval.

We use Daybreak Blue for ordinary defensive work. We use Daybreak Red with GPT-5.6-Cyber for new advanced deployments. We retain GPT-5.5-Cyber only where an existing entitlement is required.

Development line¶

• 2026-06-23 — OpenAI Daybreak security communication linked to GPT-5.5-Cyber. On 2026-06-23, the GPT-5.5-Cyber thread linked to an official OpenAI Daybreak page. The link points to an official security page in the project history. The technical claims have not been researched.

What changed¶

• 2026-05-07 — GPT-5.5-Cyber entered limited preview for critical-infrastructure defenders. It reduced refusals rather than adding material cyber capability over GPT-5.5.
• 2026-06-23 — the full GPT-5.5-Cyber release reached trusted defenders in limited release. It added stronger vulnerability-finding and patching performance.
• 2026-08-10 — Daybreak Red introduced GPT-5.6-Cyber as the current specialized model for advanced authorized work. GPT-5.5-Cyber moved to legacy access terms.

How to use this¶

As of 2026-06-23, we treat this communication as a review cue for GPT-5.5-Cyber. We do not change implementation or deployment practice until we research and verify the linked source.

1. Define an internal security scope covering systems you own or hold explicit permission to test. Request Daybreak access as an individual or organization. — https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
2. Start ordinary AppSec, incident-response, malware-analysis, and patch-validation work with Daybreak Blue rather than the legacy model. — https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/
3. Obtain separate Daybreak Red approval for authorized exploit development, red teaming, or penetration testing. Use gpt-daybreak-red-latest or gpt-5.6-cyber through the approved API project. — https://developers.openai.com/api/docs/models/gpt-daybreak-red-latest
4. Test existing GPT-5.5-Cyber entitlements with a direct Responses API request using gpt-5.5-cyber-preview. Model-listing visibility does not prove access. — https://help.openai.com/en/articles/20001259-trusted-access-for-cyber-common-issues-and-troubleshooting

Best practices¶

• Keep Trusted Access inside an internal workspace. Never route customer-facing, downstream, or third-party traffic through it. — https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
• Use Daybreak Blue as the default tier. Escalate to Daybreak Red only for advanced authorized work that requires specialized capability. — https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/
• Run security workflows in isolated environments. Review elevated tool actions with auto-review, and keep human oversight on high-risk tasks. — https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/
• Treat access as scoped and safeguards as active. Approval does not remove all refusals, and it never authorizes testing systems without explicit permission. — https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview

Superseded by this¶

• 2026-08-10 — GPT-5.5-Cyber is no longer the current specialist choice for new advanced cyber deployments. GPT-5.6-Cyber is the Daybreak Red model.
• 2026-09-05 — Daybreak Blue with GPT-5.6 Sol supersedes earlier guidance starting teams on GPT-5.5 with Trusted Access. Existing GPT-5.5-Cyber approvals stay under their original terms and do not move to Daybreak Red.

Still unknown¶

• The event record is dated 2026-06-23, while the linked first-party announcement is dated 2026-06-22. We retain the event date as 2026-06-23 and track the source date separately.
• OpenAI labels GPT-5.5-Cyber legacy in current support guidance. They publish no universal retirement date, and they do not guarantee continued availability for previously approved accounts.

Sources¶

Agent brief¶

• Subject: project:gpt-5-5-cyber, thread gpt-5-5-cyber, 1 dated events 2026-06-23 → 2026-06-23.
• Practical note: As of 2026-06-23, treat this security-related official communication as a review cue for GPT-5.5-Cyber; do not change implementation or deployment practice until the linked source is researched and verified.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.