OpenAI's GPT-5.5 Cyber AI Beats Anthropic's Banned Mythos Mannequin—And No one's Shutting It Down – Decrypt

OpenAI’s cybersecurity mannequin simply beat the Anthropic Mythos AI mannequin that the U.S. authorities yanked offline—and it is nonetheless up and working.

On June 22, OpenAI introduced the complete launch of GPT-5.5-Cyber as a part of its Dawn cyber protection program. On CyberGym—a benchmark developed at UC Berkeley that presents AI brokers with 1,507 recognized software program vulnerabilities from 188 open-source tasks and scores them on what number of they will reproduce in a managed surroundings—the up to date mannequin reached 85.6%.

Anthropic’s Mythos 5 sits at 83.8% on the identical leaderboard. Claude Opus 4.7, Anthropic’s extra broadly obtainable mannequin, scored 73.1%.

A less-than-two-point hole on any benchmark would usually be unremarkable. The context right here will not be. Anthropic’s Mythos 5 and Fable 5 had been pulled offline on June 12 after the Donald Trump administration issued an emergency export management directive citing nationwide safety.

The federal government pointed to a jailbreak—a method for bypassing an AI mannequin’s built-in security limits, much like discovering a grasp key that unlocks a high-security door. Anthropic had no dependable technique to confirm consumer nationality at scale, so it disabled each fashions for everybody, all over the place.

A number of the injury was self-inflicted. Anthropic spent months describing Mythos as one of the crucial succesful—and most harmful—AI fashions ever constructed, warning in its personal launch documentation that its cybersecurity skills might trigger critical hurt with out the fitting restrictions. Anthropic CEO Dario Amodei revealed an essay on June 10 evaluating frontier AI fashions to plane that security regulators ought to have the ability to floor in the event that they fail audits.

A number of days later, the federal government grounded Anthropic’s plane.

It wasn’t the one alarm that week. Anthropic had already come underneath hearth over a hidden filter in Fable 5 that silently degraded the mannequin’s outputs for customers it suspected of constructing competing AI—with out telling them—and was pressured to apologize and reverse the coverage.

A special playbook

Whereas Anthropic negotiates with the Commerce Division and continues its lawsuit towards the Trump administration, OpenAI is extending its attain. Dawn has signed cybersecurity partnerships with Australia, Canada, France, Germany, Japan, South Korea, and EU establishments together with the European Union Company for Cybersecurity.

Twenty-eight safety companies—together with CrowdStrike, Cisco, and Cloudflare—have joined its Cyber Associate Program to embed GPT-5.5 into their merchandise for vetted prospects. Per OpenAI’s personal weblog, Codex Safety device has scanned over 30 million commits throughout 30,000 codebases and logged greater than 500,000 fastened vulnerabilities since launching in March.

The corporate can be increasing a associate program so safety companies can combine these capabilities into their very own instruments, and it launched “Patch the Planet,” an initiative to assist repair vulnerabilities in broadly used open-source tasks.

That stated, GPT-5.5-Cyber will not be for basic use. It is obtainable solely to verified safety professionals, and OpenAI ran pre-deployment checks with federal businesses—together with the Heart for AI Requirements and Innovation and the Workplace of the Nationwide Cyber Director—earlier than launch. That is the identical restricted-access strategy Anthropic tried with Mythos, however OpenAI cleared its strategy with the federal government first.

As of June 23, Fable 5 and Mythos 5 stay offline—eleven days right into a suspension with no official restoration date from Anthropic or the Commerce Division.