OpenAI has recently unveiled its latest cybersecurity model, following the release of Anthropic’s powerful Cloud Mythos model, which is capable of identifying critical vulnerabilities in operating systems and web browsers.
Like Cloud Mythos, OpenAI’s newly released model will be available to a limited number of partners with the aim of preventing misuse of powerful AI models.
GPT‑5.4‑Cyber
This model, named GPT‑5.4‑Cyber, is specifically designed for defensive security operations. Equipped with a permissive nature, the model has a “lower denial threshold”, highlighting that it will not block sensitive security requests that normal models might consider risky.
When it comes to features, the program includes “thousands of verified individual protectors and hundreds of teams responsible for protecting critical software.”
It includes new features such as binary reverse engineering, which allows defenders to analyze compiled software for malware and vulnerabilities without needing the original source code.
Who will get access?
OpenAI’s model will be available to the “highest tier” of people and organizations, including security vendors, organizations and researchers vetted in its Trusted Access for Cyber (TAC) scheme.
Higher-level access may require users to waive “zero-data retention” (ZDR) so that OpenAI can maintain visibility into how models are being used.
“This is especially true for developers and organizations accessing our models through third-party platforms, where OpenAI may have less direct visibility into the user, environment, or intent of the request,” OpenAI writes in its blogpost.
The launch of the GPT‑5.4‑Cyber Model is based on some guiding principles. The first principle is based on democratizing access, which makes tools available to legitimate actors of all sizes through objective criteria.
OpenAI also aims to build ecosystem resilience by supporting the broader community through grants such as the $10 million program and open-source initiatives like Codex Security.
future outlook
OpenAI classifies GPT‑5.4 as a “high” cyber capability model under its readiness framework.
OpenAI wrote, “We believe that the class of safeguards in use today adequately reduces cyber risk to support widespread deployment of existing models. We expect that versions of these safeguards will be sufficient for upcoming more powerful models.”
