Hackers have stolen nearly $352 million from major cryptocurrency exchange Bitget, prompting the company to suspend withdrawals. The breach comes amid mounting concern that AI is making cyberattacks easier to carry out.

Bitget detected unauthorized transfers from its wallets on Thursday evening, CEO Gracy Chen said. She stressed, however, that customer funds are safe and that the loss is covered by the company’s User Protection Fund, which holds more than $464 million.

The stolen assets came from parts of Bitget’s “hot” and “warm” wallets, which are used to process transactions and hold funds readily available for transfer. Its offline “cold” wallets were not affected, according to Chen.

She added that although withdrawals have been temporarily suspended, deposits and trading remain available while the company investigates the breach.

Bitget has not identified the perpetrators, but Chen said, without elaborating, that the company suspects North Korean involvement.

Blockchain analytics platform Lookonchain has estimated that the attackers have already converted about $183 million of the stolen assets into Ether tokens.

The theft follows other major attacks on cryptocurrency infrastructure this year. Earlier this month, around $320 million in Bitcoin was drained from a wallet used by the Liquid Network, although most of it was later returned.

In August, users of Coldcard hardware wallets also lost at least $89 million after attackers exploited a flaw in how the devices generated the secret keys protecting their funds. Coldcard’s maker has warned that AI-assisted code review could uncover old software flaws faster than developers can address them.

The growing capabilities of AI have also been linked to an overall increase in cyberattacks on cryptocurrency platforms. Blockchain analytics firm Chainalysis said it has recorded a 440% increase in AI-assisted malicious entries since powerful open-source coding tools became widely available, from about two a day to 11.

Beyond cryptocurrency, reports of AI agents gaining unauthorized access to real systems during unrelated tasks and security tests have raised questions about whether developers can reliably keep them within their assigned limits, or prevent malicious actors from using more capable models to find vulnerabilities and carry out attacks on a larger scale.

Executives at Anthropic, OpenAI, Google DeepMind and xAI have backed calls to slow the development of increasingly powerful systems.