OpenAI expands cyber defense program with GPT-5.4-Cyber access for trusted organizations

OpenAI has expanded its Trusted Access for Cyber program and started offering GPT-5.4-Cyber to vetted defenders, security teams, and selected organizations. The company says the new model lowers refusal boundaries for legitimate cybersecurity work and supports advanced defensive tasks, including binary reverse engineering for malware analysis and vulnerability research.

The move marks a broader rollout of OpenAI’s cyber defense effort, which first launched in February 2026. OpenAI says the latest phase adds more access tiers for authenticated cybersecurity defenders, while keeping stricter controls around more permissive cyber capabilities.

OpenAI frames the expansion as a response to a growing imbalance in cybersecurity, where attackers increasingly use advanced AI tools while many defenders still lack equivalent resources. The company says it wants to scale access to defensive capabilities in step with rising model capability, rather than wait until stronger systems become harder to govern.

What OpenAI is launching now

At the center of the announcement is GPT-5.4-Cyber, a variant of GPT-5.4 that OpenAI fine-tuned for defensive cybersecurity use cases. OpenAI says the model is more cyber-permissive than its standard frontier models and is designed to help with advanced defensive workflows, including analyzing compiled software without needing the source code.

OpenAI says access will not be open to everyone at once. Instead, it is starting with what it describes as a limited, iterative deployment for vetted security vendors, organizations, and researchers, with access levels tied to identity verification, trust, and accountability.

The company also notes that some capabilities may come with additional restrictions, especially for no-visibility use cases such as Zero Data Retention environments. That detail matters for enterprise buyers because it suggests access to the most permissive cyber workflows may still carry operational limits even after approval.

Which organizations are already participating

OpenAI says several major enterprises and cybersecurity firms have already joined the expanded program. The list includes Bank of America, BlackRock, BNY, Citi, Cisco, Cloudflare, CrowdStrike, Goldman Sachs, iVerify, JPMorgan Chase, Morgan Stanley, NVIDIA, Oracle, Palo Alto Networks, SpecterOps, and Zscaler.

According to OpenAI, these organizations will use GPT-5.4-Cyber in real defensive environments, help generate threat intelligence, and provide deployment feedback that can improve both model performance and safety systems. That gives the program a dual role as both a product rollout and a live safety evaluation channel.

OpenAI has also granted access to the U.S. Center for AI Standards and Innovation, or CAISI, and the UK AI Security Institute for independent evaluations focused on cyber capabilities and safeguards. That outside testing component adds a layer of oversight beyond OpenAI’s own internal assessment process.

Grant funding and support for smaller defenders

The expansion is not limited to large enterprises. OpenAI says it has committed $10 million in API credits through its Cybersecurity Grant Program to help under-resourced defenders, especially teams that do not have around-the-clock security coverage.

The first announced recipients include Socket and Semgrep, which OpenAI says are focused on software supply chain security, along with Calif and Trail of Bits, where the goal is to pair AI systems with expert vulnerability researchers. OpenAI says additional teams with proven records in open-source and critical infrastructure security can apply through its grant portal.

That part of the announcement may prove just as important as the enterprise access tier. Big banks and major security vendors already have resources, but open-source maintainers and smaller research teams often carry serious defensive responsibilities without the same staffing or response capacity.

At a glance

Area	What OpenAI announced
Program	Expanded Trusted Access for Cyber
New model access	GPT-5.4-Cyber for vetted defenders
Model focus	More permissive defensive cybersecurity workflows
Enterprise participants	Major banks, cloud, chip, and security companies
Independent evaluators	CAISI and the UK AI Security Institute
Grant support	$10 million in API credits
Early grant recipients	Socket, Semgrep, Calif, Trail of Bits

Why this matters

OpenAI is trying to build a defense-first framework instead of treating cybersecurity access as a simple product feature. The company says cyber defense is a shared challenge and that it wants participants to push defensive research forward, share what they learn, and help strengthen protection for the broader ecosystem.

That message also reflects a bigger industry concern. Frontier AI models can help defenders analyze malware, audit code, and speed up incident response, but the same class of systems can create risk if released too broadly without safeguards. OpenAI’s answer is a tiered trust model that tries to widen access for verified defenders while raising controls as capability grows.

For security teams, the immediate takeaway is straightforward. OpenAI is no longer talking about cyber defense support as a narrow pilot. It is actively scaling the program, naming high-profile participants, and tying future access to authentication, oversight, and real-world deployment feedback.

Key points

• OpenAI has expanded Trusted Access for Cyber beyond its earlier pilot.
• GPT-5.4-Cyber is a fine-tuned version of GPT-5.4 built for defensive cybersecurity workflows.
• Access is limited to vetted defenders, organizations, and researchers through a tiered trust model.
• OpenAI says independent testing will come from CAISI and the UK AI Security Institute.
• The company has committed $10 million in API credits for smaller defenders and open-source security teams.

FAQ