Cryptocurrency exchanges Binance and Coinbase are in talks with AI startup Anthropic for access to the Claude Mythos Preview model, according to The Information.
The project released a new neural network in early April, but it is not publicly accessible. Participants in Project Glasswing—AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, and others—will be able to test Mythos in secure conditions.
The restrictions stem from the model's claimed capabilities. According to the developers, it can scan millions of lines of code and identify vulnerabilities in seconds. During testing, the neural network uncovered a 27-year-old vulnerability in OpenBSD and a 16-year-old bug in FFmpeg.
Research from Anthropic has shown that AI agents based on Mythos can autonomously hack smart contracts, simulating damages in the millions of dollars.
Traditional audits take weeks, while the new model finds effective attack vectors in seconds.
“Mythos and future models will enable even deeper testing of software and systems at scale. This will accelerate both digital threats and digital defenses,” said Coinbase's Chief Security Officer Philip Martin, confirming the discussions with Anthropic.
Binance's Chief Security Officer Jimmy Su also emphasized that the exchange is exploring “how AI advancements can strengthen cybersecurity while simultaneously creating new risks.” He stated that as part of this effort, the platform is experimenting with AI to more quickly identify vulnerabilities in its internal systems.
Clear Street analyst Owen Lau noted that the threat posed by Mythos is still not entirely clear. He believes this will create a negative narrative for cryptocurrency exchanges, but only temporarily.
“In the long run, these platforms will be among the first to defend against such AI agents,” he remarked.
It is worth noting that regulators worldwide have expressed concerns about Mythos's capabilities, with emergency talks already taking place in the UK involving the government’s cybersecurity agency and major banks.