Experts Warn of Widespread Crypto Hacks Enabled by AI Agents

Participants at the Wyoming Blockchain Symposium have raised alarms about the potential of AI agents to significantly reduce the costs and scale of attacks targeting cryptocurrency owners. They believe that automation could enable cybercriminals to simultaneously exploit vulnerabilities in wallets, passwords, and networks across numerous potential victims.

https://www.youtube.com/live/E12dLcBT2uI?si=t9GRyclsam0-9MC9

Ryan Kirkley, co-founder and CEO of Global Settlement Network, expressed concern that the industry often views autonomous agents solely as beneficial tools, underestimating their potential for misuse by attackers.

“We behave as if agents are always good. I think this is a fatal misconception almost everywhere related to these systems today,” he stated.

According to TRM Labs, the crypto sector experienced 207 hacks in the first half of 2026, the highest number recorded in any six-month period by the firm. The total damages amounted to $972 million.

Kirkley noted that previously, it was economically unfeasible for cybercriminals to invest significant time and resources to target individuals with relatively small crypto portfolios. However, AI agents can change this dynamic, allowing a single autonomous system to potentially target a vast number of victims at once.

The expert suggested that in such a scenario, even the current significant attacks on crypto protocols might seem minor compared to the overall damage caused by widespread automated campaigns.

TRM's data indicates that attacks are already becoming more prevalent. The number of incidents doubled from 83 cases in the first half of 2025 to the same period in 2026. Infrastructure and operational compromises, including the theft of private keys and seed phrases, accounted for only about 15% of incidents but resulted in approximately 76% of total losses.

Simpler and Faster

Bill Labun, Vice President of Technical Operations at the Web3 Foundation, agreed that the advantages of autonomous systems simultaneously simplify the work of criminals.

“This means there is less friction for the bad guys. For the good, the bad, and the neutral,” he noted.

The primary benefit of AI agents lies in their ability to autonomously execute sequences of actions: gathering information, interacting with external services, working with code, and utilizing available tools without constant human supervision.

These same capabilities also enable the automated search for targets and vulnerabilities.

Defenders are already leveraging such systems. In July, the Ethereum Foundation employed AI agents to analyze critical components of the blockchain. These systems examined code, sought potential vulnerabilities, and prepared materials for proof-of-concept demonstrations.

The foundation emphasized the need for manual verification of results, as a significant portion of the candidates identified by the agents turned out to be false positives, duplicates, or issues outside the scope of the study.

How Much Power to Give an Agent

Participants in the discussion also addressed the risk of connecting autonomous systems directly to users' financial instruments.

“The idea of giving one agent all this power — access to my credit card data, security information, Social Security, all personal data and various accounts — seems frightening without clear limitations,” stated Fahmi Syed, president of Midnight Foundation.

Kirkley believes that managing permissions is a solvable issue but considers the compromise of the agent itself or the environment in which it operates to be a more serious threat.

“Are we not creating a new attack vector where the agent can be compromised and the entire wallet emptied?” he questioned.

This problem is becoming more relevant as wallets and crypto services emerge that allow AI to manage assets independently. On August 6, MetaMask opened access to Agent Wallet, where users can set spending limits, permitted networks, addresses, and protocols, allowing the system to autonomously conduct transactions within set boundaries.

MetaMask also warned of the risk of "prompt injections." If an agent is analyzing data from external sources while being able to initiate financial operations, a hidden malicious instruction could potentially lead to an irreversible on-chain transaction.

To mitigate this risk, the wallet separates the decision-making by the model from policy verification and transaction signing.

Other Risks

Labun highlighted another potential issue: a false sense of privacy. Even if a blockchain or application conceals transaction contents, accompanying metadata may enable algorithms to correlate separate pieces of information and reconstruct links between users.

“This is what truly worries me: metadata leaks. People will think they are safe because they are using a private network,” remarked the Web3 Foundation representative.

In addition to security, panel participants identified a lack of trust as one of the main barriers to the widespread adoption of AI agents. Labun reminded attendees that current language models still have the capacity to generate erroneous information.

“My LLMs still sometimes hallucinate. I wouldn’t want to entrust my retirement account to their care,” he noted.

Richard Shorten, founder of Silvermine Capital Advisors, believes that the development of agent technologies is outpacing the speed at which users and companies can comprehend the implications of their implementation. He stated that the challenge lies in transforming technical capabilities into systems that people are willing to entrust with at least parts of their business or daily lives.

Who is Responsible for an Agent's Mistakes?

Another unresolved issue concerns legal accountability for autonomous actions. Kirkley questioned who should be held responsible if an agent independently breaks the law, transfers funds to the wrong recipient, or executes an irreversible transaction.

“If your agent does something illegal or spends money it shouldn’t have, how do you get it back? Who makes the final decision? The question is, who ultimately bears the responsibility?” he said.

This issue is particularly pronounced in blockchains, where a completed transaction typically cannot simply be reversed.

Previously, Anthropic identified issues of trust, deception, and collusion among multi-agent AI systems.

In August, ForkLog explored the concept of the "reverse centaur," as described by Cory Doctorow, and examined who benefits from such automation.