OpenAI on Tuesday unveiled GPT-5.4-Cyber, a variant of its flagship model fine-tuned specifically for defensive cybersecurity work. The model can reverse-engineer compiled software, find vulnerabilities in code without seeing the source, and probe systems for weaknesses that human security teams would take weeks to find.

The timing is not a coincidence. One week ago, Anthropic dropped Mythos Preview and told the world it had found security vulnerabilities in every major operating system and web browser. The announcement triggered emergency meetings at the Treasury Department, the Fed, and the Bank of England. Governments panicked. Wall Street sold off software stocks.

OpenAI watched all of that happen and said: we have one too.

A Model You Cannot Use

GPT-5.4-Cyber is not coming to your ChatGPT subscription. OpenAI is restricting access to the highest tier of its Trusted Access for Cyber program, which means verified cybersecurity defenders, vetted security vendors, and hand-picked researchers. Everyone else gets the standard GPT-5.4 with its usual guardrails.

The reason is straightforward: GPT-5.4-Cyber is purposely fine-tuned with fewer capability restrictions. In plain English, the safety guardrails are lowered on purpose. It can do things the regular model refuses to do. That includes binary reverse engineering, which lets security professionals analyze compiled software for malware, vulnerabilities, and security weaknesses without needing the original source code.

This is the kind of capability that defenders need and attackers want. The line between the two is a background check and an application form.

The Mythos Effect

Anthropic played the cybersecurity card first, and it played it masterfully. Project Glasswing, the responsible disclosure program attached to Mythos, gave Apple, Google, JPMorgan, and Microsoft 90 days to patch the vulnerabilities Mythos found. The message was clear: our AI is so powerful it can hack the entire internet, and we are the responsible ones for telling you about it.

OpenAI could not let that narrative stand unchallenged. If Anthropic owns the story that its AI is the most capable cybersecurity tool on Earth, that is worth billions in government contracts, enterprise deals, and IPO positioning. So OpenAI responded the only way it could: by releasing its own cybersecurity model and framing it as the defensive counterweight.

CNET noted the timing plainly: this is the latest chapter in the ongoing battle for dominance between OpenAI and Anthropic. Reuters, Bloomberg, and Axios all reported the same thing. Nobody is pretending this is a coincidence.

The Difference Nobody Is Talking About

Here is what matters: GPT-5.4-Cyber is a fine-tuned version of an existing model. Anthropic Mythos Preview is a fundamentally new model that Anthropic claims represents a generational leap. OpenAI took GPT-5.4 and turned down the safety dials. Anthropic built something new from scratch that it says is too dangerous to release publicly.

That distinction matters for the arms race narrative. OpenAI is retrofitting. Anthropic is building purpose-designed. One approach is faster. The other is potentially more capable. And both companies are watching the clock on GPT-6 (codenamed Spud, with Polymarket giving 78% odds of an April launch) and whatever Anthropic has planned after Mythos.

What Happens Next

The cybersecurity AI market just became a three-horse race. Anthropic has Mythos and Project Glasswing. OpenAI has GPT-5.4-Cyber and Trusted Access. Google, which joined Anthropic's industry coalition, is presumably building its own capabilities quietly.

The uncomfortable truth is that every major AI lab is now building models that can find and exploit software vulnerabilities at machine speed. They are calling it defense. They are releasing it to vetted partners. But the capability itself does not care what you call it. A model that can find every zero-day in Windows does not become less dangerous because you gave it to a cybersecurity firm instead of a hacker.

OpenAI said it is using feedback from trusted testers to improve resilience to jailbreaks and other adversarial attacks. In other words: they know someone will try to break through the guardrails. The question is not if, but when.

First reported by Reuters. Additional reporting from Bloomberg, CNET, 9to5Mac, and Axios.