Anthropic’s Mythos isn’t threatening bitcoin. The real AI risk is at crypto exchanges

There’s a saying in crypto: not your keys, not your coins. It’s typically a warning that it’s safer to keep your crypto in an offline wallet that’s harder to attack than an online exchange where you don’t have control over the keys to the wallet housing the funds.

So far, however, the complexity of learning how to self-secure and custody crypto has been prohibitive to many prospective crypto investors, which has allowed exchanges like Coinbase and Gemini Space Station – two of the earliest in the game – to thrive.

But with money now being more digital than ever and AI getting stronger, that narrative may be flipping. AI like Anthropic’s Mythos, which is built to find software vulnerabilities at extreme speed with unprecedented accuracy, could help usher a new wave of attacks on companies built on and around crypto – an industry already working to overcome a years-long reputation of hacks, scams, and exploits.

Some investors in crypto – whose universe of investable assets has quickly grown beyond coins to include ETFs and equities covering various crypto themes – can take comfort in the fact that the Bitcoin blockchain itself has never itself been hacked and has operated securely and without interruption since 2009. The threat of Mythos-like AI probably won’t change that.

“Bitcoin is fundamentally secured by cryptography and a set of shared rules,” said Yan Pritzker, chief technology officer at Swan Bitcoin. “The cryptography itself isn’t affected by AI, and the shared rules are enforced by a network of people running Bitcoin nodes all over the world. So while AI can influence how those people think in some way, it really is very difficult to modify the rules of the network without really full consensus from the network.”

Exchanges like Coinbase, Robinhood, Gemini or Bullish, on the other hand, are perhaps the most at-risk areas due to the large amounts of personal identifiable information and money they handle.

“Any other system that deals with money in a real-time basis is going to be a place that we try to look for cyber security holes,” said Cosmo Jiang, general partner at Pantera Capital. “While the threat factor exists for everyone, it’s most likely that financial services companies or exchanges are going to be the ones that are targeted first.”

Owen Lau, an analyst at Clear Street, said to consider the reputational risk AI agents pose to crypto exchanges when assessing downside risk. Specifically, he said, they can generate large volumes of scam emails and create synthetic identites, building detailed profiles that pull information from the exchanges or other retail platforms.

A double-edged sword

While AI can create new kinds of threats, the biggest exchanges argue it also presents an opportunity for them to improve security for their users. Coinbase and Binance both said they’re keen to invest in and use AI to make their platforms more secure.

“Mythos, and future models like it, will enable even deeper testing of software and systems at scale,” Philip Martin, chief security officer at Coinbase, said in a statement shared with CNBC. “This will accelerate digital threats as well as digital defense.” 

He also said that although Mythos “is a highly restricted model not available to the public, Coinbase is in close communication with Anthropic.”

Similarly, Binance’s chief security officer Jimmy Su said the company is evaluating “how advances in AI can create new opportunities to strengthen cybersecurity while also introducing new risks. As part of that work, we are experimenting with AI to help us identify vulnerabilities faster and more broadly across our systems.”

Lau said the threat of a Mythos-like AI isn’t yet clear or specific enough to make him reconsider his bullish ratings or his price targets, and he cautioned against letting short-term fear and uncertainty drive investors away from the sector.

“Near term, it will become a negative narrative for these kinds of companies,” he said of crypto exchanges. “But longer term, I would see them as one of the first batches that comes out and can protect against these AI agents.”