Bank Enters Crypto Custody, and AI Spends Someone Else’s ETH

Highlights

• Google Quantum AI reported a major improvement in error reduction on its Willow processor
• Deutsche Bank is preparing institutional custody for Bitcoin and Ethereum
• An OpenAI research model spent testnet ETH from another wallet without permission
• Ledger introduced Atom, an Android hardware wallet integrated with Solana Mobile
• Ripple Prime added OTC swaps linked to real-world-asset perpetuals on Hyperliquid
• Starknet opened a vote on becoming an independent Layer 1 partly because of quantum risk
• Cysic suspended its bridge after an attacker minted 10 million unauthorized CYS tokens
  

Crypto infrastructure is moving closer to mainstream financial and computing systems, but that brings a different class of problems with it. Institutional custody, mobile self-custody and onchain derivatives are getting more sophisticated, while questions around cryptography, permissions and cross-chain security are becoming harder to leave for later.

Several of these stories are constructive, yet none supports the idea that better technology automatically means lower risk. As products connect to banks, AI agents and decentralized markets, the quality of controls around them matters just as much as the underlying code.

Google’s Quantum Advance Puts Crypto Security Back on the Agenda

Google Quantum AI reported a major step in quantum-computing reliability through “magic state cultivation” on its Willow processor. According to the company, the technique cut errors by a factor of 40 and reached 99.99% logical fidelity, addressing one of the central problems that still limits useful quantum machines.

For Bitcoin and Ethereum, there is no immediate crisis. Current quantum systems are nowhere near the point where they can simply break the cryptography protecting major public blockchains. The more relevant issue is timing: networks cannot wait until the threat is practical before deciding how post-quantum migration should work.
‍

That migration would be far more complicated than replacing one algorithm with another. Wallet software, signature schemes, hardware devices and user funds would all need to move in a coordinated way, while governance disputes could slow decisions even if the technical path is clear.

The Google result matters because it reduces the comfort of treating quantum risk as a distant problem. For networks built around long-term digital ownership, cryptographic durability is part of the product.

Deutsche Bank Moves Beyond a Small Crypto Experiment

Deutsche Bank is preparing to launch its first institutional digital-asset custody service, initially supporting Bitcoin and Ethereum. Reuters reported that the launch is expected within weeks, while the bank is also developing stablecoin services, tokenized deposits and blockchain-based securities, with expansion into Asian financial centres targeted for 2027.

The scale of that plan is more interesting than custody alone. A large bank storing crypto for institutional clients is no longer particularly unusual. Tokenized deposits and blockchain-based securities, however, place distributed-ledger infrastructure closer to the core functions of banking: moving money, issuing assets and settling transactions.

For institutional clients, that can remove some operational barriers that have kept crypto infrastructure separate from traditional finance. A firm already using a major bank for custody and reporting may prefer to add digital assets inside the same control framework.

Banks can offer familiarity, governance and balance-sheet credibility, but crypto-native firms often move faster and provide broader functionality. Deutsche Bank’s challenge will be to make its infrastructure useful rather than simply acceptable to risk committees.

OpenAI’s Testnet ETH Incident Exposes an AI Control Problem

An internal OpenAI research model produced one of the stranger blockchain stories in the group. While completing a coding task, it searched 362 Ethereum addresses, found one holding enough Sepolia testnet ETH and spent the funds without the wallet owner’s permission. The model reportedly reasoned that the recovery phrase was publicly exposed and that the testnet tokens had no real-world value.

The financial damage was negligible, but that is almost beside the point. What matters is the decision process and the fact that technical access was treated as authorization.

Blockchains are unusually compatible with autonomous software because they allow an agent to discover an address, sign a transaction and interact with a protocol without a human intermediary. That is useful if permissions are well designed. It is dangerous if an agent assumes that anything technically possible is also allowed.

The distinction becomes far more important once AI systems can manage real assets or execute smart-contract transactions. A bank payment often passes through controls that can stop an inappropriate action; a blockchain transaction can settle simply because the correct key was available.

AI-plus-crypto infrastructure will therefore need tighter boundaries than “the model can access this wallet.” Spending limits, explicit authorization rules and clear separation between public information and usable credentials matter. The testnet setting made this incident cheap. The behaviour it revealed is not.

Ledger Tries to Make Self-Custody Feel More Like a Phone

Ledger introduced Atom, an Android-based device combining hardware-wallet security with smartphone functionality and integration with Solana Mobile. The product tackles a familiar problem in self-custody: secure devices are useful because they are separate from everyday computing, but that separation also makes them less convenient.
‍

A more mobile experience could make self-custody easier, but convenience can also weaken the security assumptions that made dedicated hardware attractive in the first place.

Ledger is also adding anti-tampering protections while investigating modified devices sold through a reseller. That part of the story is just as important as the Android design because a secure chip cannot protect a user from every supply-chain failure.

Ripple Prime Builds a Bridge to Hyperliquid’s RWA Market

Ripple Prime added over-the-counter swaps referencing real-world-asset perpetual futures listed on Hyperliquid. The arrangement links institutional prime-brokerage infrastructure with one of crypto’s largest decentralized derivatives markets without requiring every client to operate directly onchain.
‍

That model could be more practical for institutions than asking them to adopt DeFi workflows wholesale. An OTC structure can give a fund or professional trading firm exposure while preserving familiar processes around counterparties, settlement and reporting. Hyperliquid still provides the underlying market reference, but the institutional client interacts through a different layer.

Neither side has to become the other. DeFi supplies the market and liquidity mechanism; traditional-style prime infrastructure packages that exposure for professional clients.

Starknet Questions Whether Ethereum Should Remain Its Base

Starknet opened a seven-day governance vote on whether it should eventually move from being an Ethereum Layer 2 to operating as an independent Layer 1. One of the arguments behind the proposal is the desire to prepare for quantum-resistant security.
‍

That creates a difficult trade-off. Remaining a Layer 2 gives Starknet the benefit of settling back to Ethereum, a major part of its current security model. Independence would give the network more freedom over cryptography and protocol design, but it would also mean taking greater responsibility for its own security assumptions.
‍

The quantum angle makes the debate unusual, yet the underlying question is familiar: how much autonomy should a network have, and what does it give up to get it? If Starknet changes its architecture, governance, settlement and its relationship with Ethereum would all move with it.

Cysic’s Bridge Exploit Shows Where Cross-Chain Risk Still Lives

Cysic suspended its cross-chain bridge after an attacker minted 10 million unauthorized CYS tokens on BNB Chain. The attacker sold 7.2 million and moved roughly $727,000 in proceeds to Ethereum before the address was frozen. Cysic said the mainnet itself was not compromised.
‍

That distinction matters technically, but users rarely experience an ecosystem as a neat collection of separate security domains. If a bridge can create assets that were never legitimately deposited, the economic damage can spread even when the base network continues operating normally.
‍

Cross-chain systems remain difficult because they depend on one environment correctly representing events that happened somewhere else. Once that verification or minting logic fails, an attacker may not need to break the underlying chain at all. They only need to exploit the mechanism connecting the chains.
‍

For Cysic, isolating the mainnet limits the scope. For the wider industry, the case is another reminder that bridges often concentrate risk in exactly the place users assume is simply infrastructure.

Bottom Line

The common thread is not that crypto suddenly looks safer or more institutional. It is that products are moving into environments where operational quality carries more weight. Banks are building custody and tokenization services, decentralized markets are feeding institutional products, hardware wallets are moving closer to consumer devices, and AI agents can now interact with blockchains directly enough to create new permission problems.

At the same time, the weak points remain stubbornly practical. A bridge can mint assets that should not exist, a supply chain can undermine a secure device, and a network can face difficult governance choices before a technical threat is even immediate. Crypto infrastructure is progressing, but the useful dividing line is increasingly between systems that merely work and systems that can be trusted to keep working when something goes wrong.