Anthropic and OpenAI Just Rewrote the Cybersecurity Playbook

Anthropic introduced Claude Mythos last week through Project Glasswing, a restricted program capped at roughly 40 organizations, including Amazon, Apple, Microsoft and JPMorgan Chase.

OpenAI followed on Tuesday (April 14) with GPT-5.4-Cyber, deploying its system to thousands of verified defenders through its Trusted Access for Cyber program. Both models can find and exploit software vulnerabilities at a scale no human team can match. What divides them is a fundamental disagreement about what to do with that power.

A Model Built to Work Without Supervision

Anthropic’s Mythos doesn’t assist security teams. It works independently. Given a target and a prompt asking it to find a vulnerability, the model reads code, forms hypotheses, tests them against a running environment and produces a complete exploit without further human input.

Anthropic confirmed that these capabilities weren’t explicitly trained into the model. They emerged as a downstream consequence of general improvements in code, reasoning and autonomy. The same improvements that make the model more effective at patching vulnerabilities also make it more effective at exploiting them.

Mythos was able to find serious security weaknesses that had been hiding in widely used software for years. Some of these flaws had gone unnoticed for over a decade, despite being reviewed many times by experts and existing tools. What stands out is that the AI model found them on its own after a simple prompt, without any ongoing human help.

VentureBeat noted that Anthropic engineers with no formal security training asked Mythos to find remote code execution vulnerabilities overnight and woke up to a complete working exploit by morning.

On a standardized security test built around real vulnerabilities in Mozilla Firefox, Mythos successfully turned known weaknesses into working exploits 181 times, compared to just two successful attempts by the earlier model. That’s a dramatic leap in its ability to both find and act on software flaws. According to Anthropic, that gap drove Anthropic’s decision to keep the model out of general circulation.

Reuters found that the model’s coding ability has given it a potentially unprecedented capacity to identify vulnerabilities and devise ways to exploit them, with the timeline for finding and fixing flaws collapsing from months to seconds.

PYMNTS reported that Project Glasswing’s partners include cybersecurity firms and infrastructure players, giving them a head start to rewrite insecure legacy code before criminals can act.

Where OpenAI’s Design Differs

GPT-5.4-Cyber is built around a different premise. Rather than autonomous operation, it’s designed to remove the friction that security professionals hit when using standard AI tools.

Axios reported that OpenAI designed the model after some cyber partners said earlier GPT models sometimes refused dual-use security queries outright. The model lets analysts examine compiled software for weaknesses without access to the underlying source code, work that previously required specialized researchers.

It’s a bet on a different theory of control. SiliconAngle noted that OpenAI shifted away from restricting what models can do and toward verifying who gets access to the most sensitive capabilities. The Trusted Access for Cyber program launched in February alongside a $10 million cybersecurity grant program and now carries tiered verification levels, with higher tiers unlocking more capable tools.

The Hacker News detailed that OpenAI expanded access to thousands of authenticated individual defenders and hundreds of teams responsible for securing critical software. Its Codex Security product contributed to fixes on more than 3,000 critical and high-severity vulnerabilities since launch.

AI Arms Race

The two positions reflect a strategic disagreement. Anthropic concluded Mythos was too capable to distribute widely, regardless of who was asking. OpenAI concluded that wider access to properly verified defenders produces better outcomes than scarcity.

Financial institutions face a real test. Reuters found that banks are particularly exposed because they run technology stacks spanning both new and decades-old systems, house undiscovered vulnerabilities and are closely interconnected.

Costin Raiu, co-founder of cybersecurity firm TLPBLACK, told Reuters that a model like Mythos would have “a field day” finding exploits in certain IBM systems, pointing to legacy technologies powering the financial industry as a prime example.

For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and Digital Transformation Newsletters.