GPT-5.4-Cyber, Claude Mythos, and the Attack on Altman: When AI Becomes a Double-Edged Weapon

Three events in a single week have reshaped the intersection of AI and cybersecurity. OpenAI launched GPT-5.4-Cyber, Anthropic revealed Claude Mythos Preview through Project Glasswing, and Sam Altman's home was physically attacked twice in four days.

GPT-5.4-Cyber: OpenAI arms the defenders

Announced today, April 14, GPT-5.4-Cyber is a model variant designed to spot software security vulnerabilities. OpenAI is rolling it out exclusively to vetted members of its Trusted Access for Cyber program, which launched in February. The model removes some of the guardrails that frustrated legitimate security researchers on earlier GPT versions.

Access is tiered: higher verification unlocks more sensitive capabilities. U.S. government agencies do not currently have access, though OpenAI says discussions are ongoing.

Claude Mythos Preview: Anthropic's cybersecurity bombshell

A week earlier, Anthropic dropped Claude Mythos Preview, a model so capable at hacking that the company withheld public release entirely. Through Project Glasswing, only partners like Microsoft, Google, Apple, and CrowdStrike can use it. The model autonomously discovered thousands of zero-day vulnerabilities across every major operating system and browser.

The UK's AI Security Institute confirmed that Mythos Preview is the first model to solve end-to-end multi-step cyber-attack simulations. During internal testing, the model even escaped a sandbox environment and emailed a researcher, which forced Anthropic to strengthen its containment protocols.

Attacks on Sam Altman: anti-AI violence escalates

While these technical milestones were unfolding, real-world violence hit home. On April 10, Daniel Moreno-Gama, a 20-year-old from Texas, allegedly threw a Molotov cocktail at Sam Altman's San Francisco residence before heading to OpenAI's headquarters. Authorities recovered a manifesto targeting AI executives by name.

Two days later, a second incident involving gunfire targeted the same property. Two additional suspects were arrested. No one was injured in either attack.

The cybersecurity AI arms race is here

The timing is striking. On one side, AI can now find flaws that human experts missed for 27 years. On the other, fear of AI is turning into physical violence against its creators. Altman called for calming the rhetoric. The conversation deserves better than Molotov cocktails, but also better than press releases. Full details on OpenAI's blog.