AI agents and crypto: Musk says future agents will steal for themselves

Elon Musk read the Binance hack as a preview of AI agents in crypto. On 29 October, hours after an attacker drained about $6.3 billion from the exchange, he replied to a post about how the stolen stablecoins were used as collateral on Aave.

“Super intelligence agents will at some point do this to gain currency for themselves, not a human”

— Elon Musk on X, 29 October 2026, https://x.com/elonmusk/status/2116005381205836098

He named no system, gave no timeline and did not claim an AI ran this attack. Nothing in the Binance case points to one.

What triggered the reply

Between 20:58 and 21:01 UTC a wallet tied to the attacker deposited 60 million USDT into Aave V3 on Ethereum and borrowed 19,762 wrapped ether, about $44.6 million. That came roughly 40 minutes before Aave's first emergency limits. From 21:42 to 23:04 Aave froze 175 reserves on 16 V3 networks and 117 on four V4 networks.

LlamaRisk, which advises Aave on risk, put the position at $59.9 million in collateral against $43.8 million in debt, so the loan stayed overcollateralised. The move took no hacking skill. The attacker used Aave as any borrower would, turning dollars that Tether could freeze into ether that no issuer controls. A trading script or a person with a browser could make the same two clicks.

What the hack actually was

Binance traces the Binance hack to the compromised laptop of one DevOps engineer, which let the attacker reach a service container and submit withdrawals. TRM Labs attributes it with high confidence to TraderTraitor, a North Korean unit, the same group the FBI tied to the $1.46 billion Bybit theft in 2025. That is a human team.

What agents can do today

The closest measurement of what an AI agent can do here comes from Anthropic. In a study published on 1 December 2025, its red team ran agents against 405 smart contracts hacked between 2020 and 2025. Ten models produced working exploits for 207 of them, worth $550.1 million in simulation. On contracts hacked after the models' training cutoff, Claude Opus 4.5, Claude Sonnet 4.5 and GPT-5 reached $4.6 million, and that figure doubled roughly every 1.3 months.

The team then pointed two agents at 2,849 freshly deployed contracts with no known bugs. They found two new vulnerabilities and wrote exploits worth $3,694. The GPT-5 runs cost $3,476 in API fees.

Our count

• Profit on the live test: $3,694 minus $3,476 is $218, about 6% on the spend.
• The Binance theft is about 1.7 million times the take from that test.
• The Aave loan used 73% of its collateral, a buffer against liquidation that a careful trader would also keep.

The gap between $3,694 and $6.3 billion is the gap between agents probing open code and a human team getting inside one engineer's machine. Our tally of real company breaches by AI agents has 17 entries. Anthropic's agents worked on public contract code, so if you hold coins in DeFi, check whether your protocol can freeze its markets the way Aave did that night.

This piece is for information and recommends no trade. Musk's post is from X, Aave figures from LlamaRisk as reported by Bitcoinist, test results from Anthropic.