OpenAI launches new model for cybersecurity teams a month after Anthropic debuts on Mythos

OpenAI Unveils GPT‑5.5‑Cyber Following Uproar Around Anthropic Myths
It’s a modest update focused on permissive cybersecurity tasks like vulnerability classification and malware analysis.
Access is limited to computers vetted in the Trusted Access for Cyber (TAC) program, unlike Anthropic’s more restricted Mythos Preview.

OpenAI has unveiled GPT-5.5-Cyber, an updated cybersecurity model that seeks to take some of the shine off Anthropic’s Mythos Preview version.

Less than a month after the release of GPT-5.4-Cyber, this is not a major update by any means and users should not expect many changes, OpenAI explained.

Instead, users should expect a trained model to be more permissive when it comes to cybersecurity tasks, making it easier to use for things like vulnerability identification, classification, patch validation, and malware analysis.

Compete with myths

“GPT-5.5-Cyber allows a smaller group of partners to study advanced workflows where specialized access behavior may be important,” OpenAI said in a blog post.

“The cyber defense ecosystem is broad, and GPT-5.5 and GPT-5.5-Cyber play different roles in meeting the needs of organizations and researchers, depending on the task, environment, and safeguards around how the model is used. For most teams, GPT-5.5 with TAC is our most robust and widely useful model for legitimate defensive work, with strong safeguards against misuse.”

As with the previous version, this edition will only be delivered to vetted cybersecurity teams. However, unlike its main competitor, Mythos, which was delivered to only a handful of companies, OpenAI’s model will be offered to a broader set of users, members of the Trusted Access for Cyber (TAC) program.

When it introduced GPT-5.4-Cyber, OpenAI said it was expanding TAC to “thousands of verified individual defenders and hundreds of teams responsible for defending critical software.”

Anthropic first revealed Project Glasswing in early April 2026, saying the Mythos Preview AI model was too powerful to be offered freely. Apparently, he was able to expose decades-old vulnerabilities in some of the most widely used operating systems out there and chain them together to create working exploits.