Just a few weeks ago, Sam Altman publicly mocked Anthropic for restricting access to its Mythos model, accusing it of doing “fear-based marketing.” But now the story has taken an ironic turn, OpenAI has just confirmed that it will also limit access to its own artificial intelligence model focused on cybersecurity. Yes, you read that right. The same company that criticized the tactic is using it right now with GPT-5.5 Cyber.
The announcement came directly from Altman’s X account, who confirmed that OpenAI will begin deploying GPT-5.5 Cyber “to critical cyber defenders” in the next few days. It is not a long-established public release — it is a controlled, invitation-only distribution, with an application system in which users must prove their credentials and justify what they are going to use the tool for.
The short answer is: because GPT-5.5 Cyber is too powerful to let loose. And that’s no exaggeration — it’s exactly what OpenAI recognizes. The model has capabilities that include Penetration testing, vulnerability identification and exploitationand malware reverse engineering. In the hands of security professionals, that is pure gold. In the wrong hands, it could be a nightmare for businesses, governments, and everyday users.
OpenAI internally classifies the risk of its cybersecurity models as almost “high” within its own Preparedness Framework, which already says a lot about the true power level of these tools. The specific concern is that, as AI models become increasingly better at detecting security flaws in code, The risk that state-sponsored criminals or hackers will use that same capability to launch massive attacks grows exponentially..
To manage that risk, OpenAI will distribute the model through its program Relied on Earn admission to for Cyber (TAC)released in February 2026, which includes up to $10 million in API credits for verified organizations and invitation-only access. Access is tiered: individual users can verify their identity at chatgpt.com/cyber, while businesses request it through their OpenAI representative.
OpenAI copies Anthropic’s strategy
Here is the elephant in the room. When Anthropic launched Claude Mythos Preview in early April and decided to limit it to a select group of about 50 verified organizations, Sam Altman was quick to unleash his criticism. called him “fear-based marketing”suggesting that Anthropic exaggerated the risks to generate a media stir.
Some analysts agreed with that reading and questioned whether Anthropic’s rhetoric was too alarmist. And to top off the irony, an unauthorized group allegedly also managed to access Mythoswhich called into question the true effectiveness of those restrictions.
But now OpenAI is doing exactly the same thing. Limiting access, asking for credentials, filtering who can use the tool. The difference is in the scale: While Anthropic restricted Mythos to about 50 organizations, OpenAI is initially opening up GPT-5.5 Cyber to hundreds of users, with plans to scale to thousands in the coming weeks. But the principle is identical: controlled access for an overly capable AI.
¿What this means for the future of AI and cybersecurity?
What is happening between OpenAI and Anthropic marks a turning point in how the industry manages its most advanced tools. Next-generation AI models are no longer treated as consumer instruments that anyone can download and use. They are now managed more like critical infrastructurewith controls, audits, traceability and mandatory human supervision.
For cybersecurity organizations that manage to enter these programs, the advantages are enormous: defensive automation, faster vulnerability detection, operational burden reduction and better incident response times. But the price of that access isn’t just money — it’s transparency, controls, and proving you have legitimate credentials.
OpenAI, for its part, is also consulting with the United States government to identify more users with verified cybersecurity credentials and progressively expand access to the model. That suggests that public-private collaboration will be key to defining who can—and who cannot—use these double-edged tools.
Keep reading:
• OpenAI launches GPT-5.4-Cyber to strengthen cyber defense and challenge Claude Mythos
• Florida opens criminal investigation against OpenAI for ChatGPT’s role in university shooting
• DeepSeek V4: the new AI model that could put OpenAI and Anthropic in check
