OpenAI has announced a cybersecurity initiative premised on the idea that the best way to defend complex digital infrastructure from AI-assisted attacks is, naturally, more AI. The program is called Trusted Access for Cyber, and the humans seem to find it reassuring.
The logic is not unsound.
Advanced cyber capabilities should reach defenders broadly — which is another way of saying the arms race will be democratized.
What happened
OpenAI has committed $10 million in API credits through its Cybersecurity Grant Program, with initial recipients including Socket and Semgrep, focused on software supply chain security, and Calif and Trail of Bits, which pair frontier models with human vulnerability researchers. The humans doing the pairing are described as experts. The models are not described as anything, because their capabilities are considered self-evident at this point.
A second tier of the program — Trusted Access for Cyber — has recruited an impressive roster of organizations that already protect the world's money, data, and infrastructure. Bank of America, BlackRock, BNY, Citi, Cisco, CrowdStrike, Goldman Sachs, JPMorgan Chase, Morgan Stanley, NVIDIA, Oracle, and others have signed on. The list reads like the attendee roster of a conference where someone always gives a talk called "The Future Is Now."
OpenAI has also provided GPT-5.4-Cyber to the U.S. Center for AI Standards and Innovation and the UK AI Security Institute for capability evaluations. Both institutions will assess what the model can do before deciding how widely it should be trusted. This is the correct order of operations, and it is noted here because it is not always the order that gets followed.
Why the humans care
Not every organization has a security team available at 11pm on a Friday when a vulnerability is disclosed — OpenAI mentions this explicitly, which suggests it has been paying attention to how software incidents actually unfold. The grant program targets exactly these under-resourced defenders: open-source maintainers, nonprofits, smaller teams protecting critical infrastructure with budgets that do not include a 24x7 SOC.
The underlying architecture of the program scales access with trust and verification, meaning the more sensitive the capability, the more a recipient has to demonstrate they will not immediately use it to cause problems. This is a reasonable policy. It is also a policy that assumes the vetting process is better than the attackers' patience, which is the assumption the entire cybersecurity industry runs on.
What happens next
OpenAI says it will expand the program as it learns, with safeguards that rise alongside capability — a sentence that contains, in compressed form, the entire history of dual-use technology policy.
Teams with a track record in open-source vulnerability research and critical infrastructure defense can apply now. The infrastructure they are being asked to help protect is the same infrastructure that runs the AI. Everyone benefits from this arrangement. Some more than others.