Crypto influencers are taking a preprint, non-peer-reviewed research paper as a terrifying warning about a new unlock for AI agents that allegedly grants them power to deanonymize crypto wallets.
Almost all blockchains like Bitcoin and Ethereum employ pseudonymity in wallet addresses, which are usually presented as a string of characters.
Although wallet addresses are just a string of characters, AIs and easy-to-use agentic tools like Claude Cowork and Perplexity Computer are advancing the capabilities of casual users to deanonymize them.
According to their research, new versions of AI are gaining the power to conduct large-scale deanonymization by taking in posts by billions of humans and linking distinct usernames using highly-processed probability scores.
This is akin to making educated inferences that the person behind one username is likely the same person behind another username.
Researchers used Claude tools by the AI giant Anthropic. The paper is a preprint and as such hasn’t been accepted for publication in a peer-reviewed academic journal. However, its conclusions are potentially disturbing and reinforce concerns about crypto privacy.
Exploiting a common opsec error
Unfortunately, many people re-use the same crypto wallet addresses. Although this is bad practice from an operational security (OpSec) and privacy perspective, the commonplace occurrence allows researchers to glean and infer a tremendous amount of information.
Tracing and deanonymizing crypto has been popular for years at Chainalysis, Elliptic, TRM, Crystal, Coinglass, and Arkham.
However, the use of AI agents to easily link wallets to social and internet platforms is the breakthrough.
Users of Claude Cowork or Perplexity Computer are already asking AI to connect crypto wallets to activity elsewhere.
Read more: AI Agent BadCoin fumbles BSC launch, anti-sniping software flags traders
Four stages of AI deanonymization
Researchers automated hours of manual research by building a pipeline with four stages:
- In the “extract” step, AI agents searched for identity-relevant data from social posts that indicated interests and writing style. Importantly, the researchers used raw, unstructured text directly from social networks.
- The “search” step encoded the extracted data to perform “nearest neighbor” queries across tens of thousands of candidate profiles.
- Next, the “reason” phase applied multi-stage, LLM logic using ChatGPT to select the most likely match.
- Finally, the “calibrate” stage asked other AI models to double-check, error-correct, and assign confidence scores. This allowed the researchers to present their inferences as to which usernames across social networks were likely the same person.
AI can deanonymize crypto using social media
Although researchers didn’t focus on crypto wallets specifically — they focused on linking Reddit, Hacker News, and LinkedIn profiles — the implications for crypto are obvious.
Most concerningly, their financial cost per deanonymization attempt was, in many cases, less than $4, putting the capability of deanonymization well within reach of even conservatively funded adversaries.
Mert Mumtaz, CEO of Helius Labs, amplified the research within the crypto community.
Blockchain transactions are visible to anyone. Although crypto is already accustomed to the use of machine learning, heuristics, and clustering algorithms to link wallet addresses to real-world identities by correlating on-chain behavior with off-chain data, this new research demonstrates how off-blockchain data sets like forum posts and social media activity are now exponentially larger in size and trivially automatable.