OpenAI expands Trusted Access and launches GPT-5.4-Cyber model

With the launch of GPT-5.4-Cyber, OpenAI is rolling out a more cyber-permissive model for vetted defenders. The company is also scaling Trusted Access for Cyber to thousands of verified individuals and hundreds of teams protecting critical software.

openai cover

TL;DR

  • TAC expanded: Scaled to thousands of verified defenders and hundreds of teams protecting critical software
  • Verification-first access: KYC, identity verification, and trust signals unlock sensitive cyber capabilities without manual approvals
  • Deployment principles: Democratized access, iterative safeguards against adversarial pressure, and ecosystem resilience via grants and open-source support
  • GPT‑5.4‑Cyber introduced: More cyber-permissive GPT‑5.4 variant for vetted defenders; lower refusal boundary for defensive workflows
  • Newly enabled task: Binary reverse engineering for analyzing compiled software for malware potential and vulnerabilities
  • Access and constraints: Limited rollout to vetted vendors/researchers; potential restrictions for Zero-Data Retention and third-party platform usage

OpenAI is expanding its Trusted Access for Cyber work as it prepares for “increasingly more capable models” in the months ahead—starting with a new cybersecurity-tuned variant of GPT‑5.4. The update scales the company’s TAC program to thousands of verified individual defenders and hundreds of teams focused on protecting critical software, while also introducing GPT‑5.4‑Cyber, a model trained to be more cyber-permissive for defensive use cases.

A “trusted access” approach built around verification and guardrails

The thrust of TAC is straightforward: keep broad access to general models, but use identity verification and trust signals to unlock more sensitive cyber capabilities without relying on ad hoc, manual approvals. OpenAI frames the overall strategy around three principles:

  • Democratized access backed by strong KYC and identity verification, intended to reduce arbitrary decision-making about legitimate use.
  • Iterative deployment, with safeguards updated as OpenAI learns more from real-world usage and adversarial pressure (including jailbreak attempts).
  • Ecosystem resilience, via grants, open-source security support, and tools designed to accelerate vulnerability discovery and remediation.

This sits alongside earlier initiatives OpenAI cites, including its Cybersecurity Grant Program, the Preparedness Framework, and Codex Security, which it says has contributed to over 3,000 critical and high fixed vulnerabilities since the recent launch.

GPT‑5.4‑Cyber: fewer capability restrictions for vetted defenders

The headline model change is GPT‑5.4‑Cyber, described as a version of GPT‑5.4 that lowers the refusal boundary for legitimate cybersecurity workflows and enables advanced defensive tasks. OpenAI specifically calls out binary reverse engineering as a newly enabled capability—useful for analyzing compiled software for malware potential, vulnerabilities, and security robustness when source code isn’t available.

Because this model is intentionally more permissive, OpenAI says access will roll out in a limited, iterative deployment to vetted security vendors, organizations, and researchers. The post also notes that access to “permissive and cyber-capable models” may include restrictions, particularly for no-visibility uses such as Zero-Data Retention, and especially when models are accessed via third-party platforms where OpenAI has less visibility into context and intent.

How access works: individuals and enterprises

OpenAI outlines two primary on-ramps into TAC:

Existing TAC customers can also express interest in additional access tiers via a Google Form link: can express interest.

What OpenAI says comes next

OpenAI argues that current safeguards are sufficient for broad deployment of today’s models, but draws a clear line between general availability and cyber-permissive variants, which it says require more restrictive deployments and appropriate controls. Longer-term, the company expects “more expansive defenses” will be needed as future models exceed today’s purpose-built cybersecurity systems.

Original source: https://openai.com/index/scaling-trusted-access-for-cyber-defense/

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community