However, the two biggest incidents were not simple smart-contract exploits of the type that could be engineered by AI.
In one, a group linked to North Korea withdrew approximately $285 million from the Drift protocol after a six-month social-engineering campaign that gave it admin access. For another, the attacker exploited a single-verifier flaw that allowed approximately $292 million to be embezzled from the Kelp DAO.
Another example emerged on Tuesday, when Humanity Protocol, a decentralized human-identification service, lost more than $30 million due to a private-key compromise. CoinDesk discovered that a hacker gained access to three out of six private keys on an employee’s laptop,
Therein lies the problem. While the most obvious smart-contract signals may be exactly what Anthropic’s filters are designed to catch, the biggest losses don’t require a contract bug.
Ledger’s Guilmette said these exploits come from familiar weak points: social engineering, poor signature flow, exposed keys and human error.
Models like Fable do not need to hand over ready-made exploits to change the economics of an attack. It can read public repositories, compare older versions of software, summarize audit reports, and draft concrete messages that look for small operational mistakes that humans miss.
“These exploits are rooted in social engineering and human error.”
In such an environment, a defender has to secure every key path, every dependency, every signature flow, and every privileged account. Because AI accelerates the scouting phase, the final signing phase becomes more important. Private keys need to be kept in a place where compromised laptops can’t access, and users need a trusted screen that shows exactly what they’re approving.