Crypto executives warn AI agents could scale wallet attacks
At the Wyoming Blockchain Symposium, panelists said autonomous agents could lower friction for attackers while leaving users with unresolved questions about control, privacy and liability.
By The Third AnglePublished 3 min read
Illustrative AI imagery; it does not depict a specific agent, wallet or attack. Photo: Unsplash · Unsplash License
AI agents could make crypto attacks faster and cheaper to scale, executives and researchers warned during a panel at the Wyoming Blockchain Symposium. The Block reported that speakers focused on a risk that is still prospective: autonomous software may reduce the effort needed to probe wallets, credentials and connected systems, even though no attack was presented as evidence at the session.
The panel coverage from The Block quoted Global Settlement Network CEO Ryan Kirkley saying, “We’re acting like agents are always good.” The event’s organizer page confirms the Aug. 17–20 conference, hosted by SALT and Kraken, and lists decentralization of AI among its themes.
Lower friction cuts both ways
Kirkley told The Block that an agent could potentially go after many targets at once, turning today’s isolated wallet or bridge incidents into a larger automation problem. Web3 Foundation vice president Bill Laboon made the same directional point more briefly: “less friction for the bad guys.” Both comments describe a possible attack-surface change, not a measured increase in losses.
The concern is not limited to signing transactions. Panelists also raised credential discovery, privacy leaks and the risk that an agent’s permissions could be widened by a malicious prompt, a compromised tool or a failure in the software controlling it.
The control problem remains unresolved
Midnight Foundation president Fahmi Syed argued that agents need explicit parameters, while the panel questioned who would be responsible if an agent were taken over and drained a wallet. Laboon also cautioned that language models can still hallucinate, a limitation that becomes more consequential when software can act on a user’s behalf.
The evidence supports a narrower conclusion than a forecast of imminent billion-dollar losses: crypto teams adopting agentic tools need least-privilege permissions, transaction simulation, human review for irreversible actions and clear recovery paths. The next useful evidence will be reproducible incident data showing whether those controls fail in deployed systems.
