What happened
BeInCrypto, in collaboration with researchers from Dune, published a Wednesday piece laying out five concrete rules for crypto holders who let AI agents touch their wallets. The outlet framed the guidance as a response to a real, growing attack surface, not a hypothetical. Agents built on top of large language models can now read balances, sign transactions, route through DEXs, and bridge assets with minimal supervision, and that convenience is the vulnerability.
The rules, per BeInCrypto's writeup, focus on how users hand over permission rather than on the model itself. The five, summarized: use a hardware wallet for anything the agent doesn't need to spend, give the agent its own isolated sub-account funded only with what it needs, set strict spending caps and allowlisted contracts, never paste a seed phrase or full private key into an AI chat interface, and treat any prompt the agent receives from an external source as hostile input.
Dune researchers contributed on-chain context on how often agent-controlled addresses interact with the same handful of DeFi venues, which makes allowlisting practical. BeInCrypto's piece sits at beincrypto. com/ai-agent-drain-crypto-wallet-rules.
Why it matters
The user is the custodian, and that hasn't changed. What has changed is that users are now delegating signing authority to software they don't fully understand. An AI agent pointed at a wallet with full transfer rights is, functionally, a hot key held by a third party.
A prompt injection attack, where a malicious website or token metadata field smuggles instructions into the agent's context, can turn a routine rebalance into an approval for a drain contract. Security firms flagged this class of attack in 2025, and the risk gets worse, not better, as agents gain tool access to bridges and lending protocols. The BeInCrypto and Dune framework matters because it's one of the first retail-focused checklists written in plain language.
