The artificial intelligence (AI) startup is expanding “Project Glasswing,” which offers select organizations early access to the AI, Bloomberg News reported Thursday (April 16).
This came after Anthropic learned of Mythos’ ability to spot — and possibly exploit — weaknesses in cyber defenses, the report added.
“That is in the very near term, in the next week,” said Pip White, Anthropic’s head for the U.K., Ireland and northern Europe, told Bloomberg Television. “As you would expect, the engagement I have had from UK CEOs in the last week has been significant.”
Earlier this month, Anthropic said it would hold off on releasing Mythos to give cybersecurity experts the opportunity to test it.
The company says testing by Mythos has uncovered thousands of “zero-day” vulnerabilities, including within every major web browser and operating system.
Advertisement: Scroll to Continue
The Glasswing program, the report added, initially included Big Tech companies like Amazon and Microsoft. After the program was announced, U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened a meeting with Wall Street banks to discuss Mythos and similar AI models.
As Bloomberg notes, the lack of knowledge about Mythos has banks and government agencies worried. Bank of England Governor Andrew Bailey said this week that regulators need to rapidly evaluate the threat the model might present.
“We are putting our own safeguards and our own limitations around this product because we know how powerful it can be,” White said.
PYMNTS wrote earlier this week about the contrast between Mythos and OpenAI’s newly released GPT-5.4-Cyber.
“Both models can find and exploit software vulnerabilities at a scale no human team can match,” that report said. “What divides them is a fundamental disagreement about what to do with that power.”
Rather than assisting security teams, the report continued, Mythos works independently. Assigned a target and given a prompt asking it to spot a vulnerability, the model reads code, forms hypotheses, tests them against a running environment and produces a complete exploit with no additional human input.
“Anthropic confirmed that these capabilities weren’t explicitly trained into the model,” the report said. “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.”
GPT-5.4-Cyber is designed around a different premise, PYMNTS added. Instead of autonomous operation, it’s built to remove the friction that security professionals encounter when employing standard AI tools.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
