Hong Kong Monetary Authority’s Project Sage targets responsible AI agentic finance
“Having just spoken on the intersection of AI agents and financial infrastructure at HK Fintech Week, it’s encouraging to see the HKMA taking agentic finance seriously with Project SAGE. As AI agents move from making recommendations to executing payments, questions around identity, interoperability and accountability become increasingly important.”
“But we also need infrastructure capable of supporting transactions at machine speed. At Espresso, we’re focused on providing high-throughput blockchain settlement with sub-second finality, giving sovereign chains the performance needed for this emerging economy. Trust, speed, and scalability need to develop together if agentic finance is to thrive.”
– Charles Lu, co-founder and COO, Espresso Systems
“Overall, this is a very good initiative in my opinion, and the focus areas all make sense. This is a welcome step because financial regulations guidance and the encouragement to try out different agentic solutions in a contained environment in a government-sponsored program is very helpful for innovation. Currently, there are a lot of gaps associated with the agentic economy that people are focusing on, but one of the key areas is regulatory clarity. So, this is a good starting point.”
– Chandler Fang, co-founder, t54 Labs
“Project SAGE’s focus on delegated authority addresses a practical challenge: users need to decide what an agent can do with their money, and platforms need to enforce those decisions. I would approach an AI agent the way a prop trading firm approaches a human trader: define its mandate, limit its capital and permitted markets, and stop trading automatically at a specified drawdown. Those controls should operate outside the model, with permission to change them kept beyond the agent’s reach.”
“A spending limit offers little protection if the agent can authorize its own increase. Users should take responsibility for the risks they choose and the access they grant. Platforms should make that access explicit, limited, and easy to revoke. At Quote.Trade, that thinking informs our approach to agent permissions and automated drawdown kill switches.”
– Chris Mack, CTO, Quote.Trade
AI agents
“AI adoption is no longer a sole differentiator for banks, as most institutions have some form of AI deployed in their operations already. The real differentiator now is data readiness. Institutions that clean, label, reconcile, govern, and continuously maintain their data will successfully scale AI deployments and discover new ways to provide value to customers.”
“In the age of personal AI agents, loyalty is no longer won by removing friction in the customer experience. Banks will retain loyal customers by utilizing AI to deliver advanced fraud prevention, seamless onboarding, proactive financial advice, and personalized incentives and rewards.”
“Human experts in banking and financial services are already taking on new roles in supervising AI agents. As agentic AI continues to evolve, reskilling and redesigning the workforce is not a one-time fix. Banks that succeed will continuously train, upskill and adapt their workforce to ensure human agent roles and responsibilities are optimized to deliver value where AI cannot.”
– Mamta Rodrigues, chief commercial officer of banking, financial services and insurance, TP
AI and explainability
“Goldman Sachs pulled Claude Code from most of its traders and bankers this week, and the reason is worth paying attention to. The bank didn’t walk away from AI. It drew a line around agents that could write and run code on machines holding client data, with no review and no information barriers in place.
“That’s the gap we built Uptiq to close. Banks and credit unions tell us the question isn’t whether AI is useful anymore; it’s whether they can stand behind it when an examiner asks how a decision was made. So, every Uptiq agent knows who is using it and what that person is allowed to see, logs every action, gives consistent answers, and works from the institution’s own policies and products. When it’s time to decide, a person makes the call.
“Goldman has 12,000 engineers to sort this out. Most community banks don’t, which is why we think governance has to come built in from day one. That’s what bank-grade means to us.”
– Snehal Fuzele, CEO and Co-founder at Uptiq
Bitcoin
“Bitcoin perpetual funding has averaged 0.93 basis points a day over the past week on CoinMarketCap’s data, the 100th percentile of the past 90 days. The flush cleared the leverage, but the longs that survived it are paying more to stay in, not less.
“US spot Bitcoin ETFs gave back $643 million in a single session on Thursday, more than the four sessions of inflows before it had put in. One day of fund selling undid a week of fund buying, and spot order flow leaned the same way.
“CoinMarketCap’s Fear & Greed Index goes into election day at 37, down from 67 a month ago, while Bitcoin is down less than eight percent over the same period. The mood has fallen a lot further than the price.
“Over the past seven sessions, Bitcoin futures saw $2.7 billion of net aggressive selling on CoinMarketCap’s data, while spot saw net buying. The leverage is leaving through the front door; the coins are staying.”
– Alice Liu, head of research, CoinMarketCap
Ethereum
“Stablecoin supply on Ethereum has fallen 3.2% in thirty days to about $141 billion on CoinMarketCap’s data. The pool of on-chain dollars that usually sits under a rebound is thinning through this drawdown, not building.”
“Decentralized-exchange volume on Ethereum is running 49% below its 30-day baseline on CoinMarketCap’s data, two days after the exchange incident sent it to three times baseline. The burst was assets being moved, not a new wave of buying.”
– Liu
Binance hack
“Crypto has made it very hard to steal the keys that move funds, so attackers now go after the people and the software around those keys that sign transactions. In this case, it seems the attacker compromised Binance’s signing system so that it would send malicious withdrawals. Bybit followed the same pattern last year, and Bitget did in September.”
– Gonçalo Magalhães, head of security, Immunefi
“One of the striking things about this breach is how little code was actually needed on-chain. The sweeper contract was only around 130 lines. That is a useful reminder that the sophistication of an attack is not always in the smart contract itself. The hard part sometimes is getting access to a system powerful enough to make very simple code devastating.”
“The key issue here is how the attackers were able to gain access to infrastructure with enough privilege to authorize withdrawals across multiple blockchains.”
“In crypto, we naturally focus heavily on protecting private keys and auditing smart contracts. But the people, infrastructure, and workflows capable of instructing those keys are part of the security perimeter too. If an attacker compromises that operational layer, they may not need to steal a key or discover an exotic smart-contract vulnerability at all.”
“That is what makes this incident particularly interesting: the on-chain execution appears relatively simple, while the real sophistication was in reaching the privileged systems behind it. The broader lesson is that access controls, internal infrastructure and transaction-authorization processes need to be secured with the same intensity as the keys themselves.”
– Simone Imbrescia, VP of engineering, product and design, Immunefi
Paraguay central bank takes over digital lender ueno
“What happened in Paraguay is a reminder that for most people the real risk isn’t a bank failing on paper. It’s waking up and not being able to touch your money. Nearly half the country lost access overnight, and the rush to move funds crashed other banks’ apps too. That’s what a single point of failure looks like.”
“ueno looked like a modern digital bank on the front end, but the problem was old-fashioned: lending to related parties and accounting that didn’t hold up. A slick app doesn’t replace sound balance sheets or independent oversight.”
“Fully reserved, regulated stablecoins are designed for exactly this scenario. When the reserves are transparent and held separately from the issuer’s own business, users can redeem or move funds without waiting on a single institution to reopen its doors. That’s why we’re seeing stablecoin adoption grow fastest in Latin America, where people have lived through bank freezes and currency shocks before.”
“The bigger question for the region is confidence. 97% of ueno Bank’s deposits were backed by Paraguay’s deposit guarantee. Paraguay had built a reputation as a stable place to invest, and a crisis like this tests that quickly. Regulators across Latin America will be looking hard at fast-growing digital lenders.”
– Bernardo Brites, co-founder and CEO, Trace Finance
Tokenized collateral
“Tokenizing an asset does nothing to lower the standard it has to meet, which is why the clearing houses and banks moving first are bringing tokenized collateral under the rulebooks and risk models they already apply.”
“Collateral only earns its keep when it can move in real time while staying segregated until the point of settlement, and an institution that has to release it from custody early to put it to work takes on the very risk custody was there to remove.”
“Traditional markets could run custody, settlement and collateral management as separate steps because assets settled on a fixed timetable, but once collateral moves around the clock those steps stop being separate at all, so custody and settlement infrastructure have to be built to work as one.”
– Richard Clark, CEO, Zodia Custody
