Google Warns AI Is Giving Lesser-Resourced Attackers Nation-State Reach
Google's Threat Intelligence Group says criminal and state-sponsored adversaries are increasingly using AI to automate and scale attacks, with one actor — TeamPCP — planning, building and executing a mass credential-harvesting campaign in under six hours using an AI coding chatbot, a prompt and a set of agent instructions.
For crypto, the sharpest line item is the DPRK-nexus actor Google tracks as Midnight Neptune, which researchers say has integrated AI across its operational lifecycle specifically to support cryptocurrency theft. PRC- and Iran-nexus groups are likewise using large language models for target profiling, localized social-engineering lures and malware development.
The report lands days after TRM Labs, announcing its Series C expansion at a $2 billion valuation, said its AI-in-Crime Adoption Index recorded a 40% year-over-year rise in criminal AI adoption in 2026. Google says it disrupts adversarial accounts and projects when identified and has deployed real-time defenses against model-cloning attempts — but concedes attackers adapt and move on.
Term Labs Post-Mortem: How 0.5 ETH Bought an $8.5M Governance Takeover
A post-mortem published on rekt.news reconstructs the August 23 governance takeover of Term Labs' vaults in unsparing detail. The core finding: the attacker's roughly $1,500 deposit wasn't exploiting a bug — it was buying a supermajority in a voting system almost nobody used.
Term's strategy vaults used Aragon TokenVoting, with voting power held in a wrapper token that users had to opt into by wrapping vault shares. Almost nobody did. Total voting-token supply on the ETH Meta Vault stood at 0.5352 tokens; the attacker's deposit of about 0.5 ETH, wrapped, delivered 90.7% of it — and all active voting power in four of five affected USDC vaults.
The formal settings — a 50% support threshold, 5% minimum participation — looked reasonable in isolation. But thresholds do not create opposition. The attacker filed a proposal, waited out the six-day voting window, then zeroed the Zodiac Delay module's cooldown and expiration, recalled capital from legitimate strategies, and routed roughly 2,843 WETH and 1.68 million USDC into an attacker-controlled strategy with an unlimited debt ceiling.
The incident's audit status is beside the point, the analysis argues: the code did what governance told it to. The lesson for protocols with token-voting control over vaults is that participation is a security parameter, and an unused timelock protects no one.
Coinsbuy: $8M Gone, Wallets Refilled, Still No Explanation
New analysis of the August 9 Coinsbuy incident documents an unusual aftermath: the exchange-linked funding addresses refilled ten of the wallets the attacker had just emptied — roughly $3.93 million across seven deposits — within about 10 to 12 hours of the drain, before the company published its only statement on the incident.
The theft itself followed a familiar path: a 5 USDT test transaction on TRON, a parallel sweep across TRON and Ethereum totaling about $8.07 million, laundering through FixedFloat and ChangeNOW, and proceeds pushed toward Monero via exchanges, according to on-chain monitors BlockWatchdog, PeckShield and Specter.
GoPlus Security assessed the pattern as consistent with a hot-wallet private key or admin-privilege compromise — an assessment, not a confirmed root cause. Coinsbuy's statement confirmed unauthorized withdrawals, reimbursement from reserves and restored operations, but the Panama-incorporated exchange has not disclosed the attack path, the affected control layer, or evidence independently establishing its remediation. Its Twitter account has been dormant since 2020.
The refill raises its own question: an operator confident its signing authority was still exposed would not ordinarily top the drained addresses back up. Nearly a month later, the public record still cannot establish whether keys were taken at all.
TrustGrade covers the security and trust ecosystem around digital assets. For verified trust data on the platforms and firms shaping it, see trustgrade.ai.