A library of snippets
xxx
We have identified industrial-scale campaigns by three AI laboratories—DeepSeek, Moonshot, and MiniMax—to illicitly extract Claude’s capabilities to improve their own models. These labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, in violation of our terms of service and regional access restrictions.
From: Detecting and preventing distillation attacks \ Anthropic.
xxx
xxx
Anthropic published a detailed forensics report this week alleging that DeepSeek, Moonshot (Kimi), and MiniMax ran coordinated campaigns to extract Claude’s capabilities through model distillation. 16 million exchanges. Roughly 24,000 fraudulent accounts.
xxx
xxx
Millions of pounds is lost every year to impersonation scams, with Nationwide’s own customer data showing they comprise 17% of reported scams. All ages are affected by criminals pretending to be their bank or building society, although Nationwide’s figures show it disproportionately impacts those over 65 years old (with 55% affected).
Call Checker complements its existing Scam Checker service, which is used by 100k people and prevents £300k a month from being lost. The new feature enables customers to instantly confirm whether the call they’re on is genuine through Nationwide’s banking app. The screen will display either “You’re on a call with Alex” or “You’re not on a call with us.”
From: Nationwide battles impersonation scams with new call checker service.
xxx
xxx
The Financial Services AI Risk Management Framework adapts existing federal guidelines on AI risks, which are generic and abstract enough to apply to any sector, into targeted advice for banks and other financial services companies.
The framework gives institutions tools including a questionnaire to help an institution determine its current AI adoption stage and a matrix of 230 control objectives to manage risks across the technology’s lifecycle.
Because the framework categorizes controls by adoption stage, banks do not have to waste resources on controls that do not (yet) apply to their operations.
Before this week, the financial services industry had available the National Institute of Standards and Technology, or NIST, AI Risk Management Framework, released in January 2023, to provide some of the guidance in this area.
Industry groups such as the Financial Services Information Sharing and Analysis Center, or FS-ISAC, have also published white papers on adversarial threats and responsible artificial intelligence principles in the past.
The AI risk framework released Thursday is “an operationalization” of the NIST framework, “specifically tailored for financial services,” according to FSSCC.
From: Treasury issues new AI risk tools for banks | American Banker.
xxx
Tom Loosemore’s excellent piece on the impact of AI on public services is well worth reading. Tom says that a lot of public services rely on friction to stay viable and depend on slow and confusing user experiences “to put off those otherwise eligible“. However, this cannot hold. From parents seeking special needs support to property owners appealing council tax bands, it is the friction of bad service design that restrains demand, not the law. I think this is a more general issue: that friction extends to changing banks and home insurance and just about everything else. His key point, that AI agents will remove that friction and be “doggedly relentless” on the citizen’s behalf, is precisely what Kirsty Rutter and I wrote about in our paper “Where are the Customers’ Bots” in the Journal of Digital Banking 8(2), p.132-140 (2023).
In that paper, we suggested that the coming paradigm shift in retail financial services does not arise from financial institutions use of AI but from their customers’ use of AI. Customers use will AI to assess offers from financial institutions and those customers will have access to AI as powerful as the instituions themsevles, because BigTech will give it to them. This will mean individuals will not be the customers, their bots willl. Given the abilities of the bot already in the market, this is hardly hyperbolic. But what is true for financial instititions will also be true for companies of all kinds and, as Tom highlights, every public sector body too.
We are moving into uncharted waters, frankly. No sane person will ever write a letter to the council appealing a planning decision themselves when their AI agent can not only do it for them, but do it far better than they could do it themselves even if they could be bothered to. Not in the future, but right now. When it comes to the public sector, it seems that complaints (and other enquiries) are rising in both volume and complexity, and generative AI is widely suspected to be a driver, alongside wider demand and austerity pressures. The citizen‑to‑state friction points that Tom refers to (benefits, housing, children’s services and so on) are generating more formal disputes, many of which now travel through digital channels. The implications are longer queues, higher processing costs, more legalistic dispute cultures and, as the Parliamentary and Health Service Ombudsman report notes, strong incentives for public bodies to deploy their own AI for triage and response. While I am not an expert on public sector processes, I can see a rapid escalation in complaints being generated by AI and then sent to publoc bodies that triage them using AI which then results in appeals being generated by AI… and so on to infinity.
Legal and HR practitioners in the UK report that AI‑drafted grievances and claims are typically much longer, more repetitive, more legalistic, harder to parse, with statutory references and case citations that may be irrelevant or, indeed, wholly fabricated. While those practitioners focus on employment and private‑sector disputes, the same patterns are already being observed in complaints about public authorities. Which takes me on to my point, which is that both public and private sector organisations will be, for the foresseable future, at a significant disadvantage in the use of the new technology. Why? It’s because they are regulated. If my bot hallucinates in a complaint about a parking ticket, so what? But if the council worker uses an AI that hallucinates in their response to me about school places, there’s a lawsuit coming (and, of course, my agents will be only to happy to file a mountain of no-win, no-fee lawsuits).
(Now, while it is surely a good thing that having AI remove barriers will improve access to redress for people who previously lacked confidence, literacy or legal support, particularly in complex domains like social care or special educational needs, the benefits will be remain uneven: digitally literate citizens with better access to tools may gain more leverage than those who are offline or have lower literacy, potentially widening existing inequalities.)
I suspect that very few organisations are ready to deal with customers who become thousands of times smarter literally over night.
xxx
A fraudster who was part of a scheme to target Tube passengers with scam text messages laundered £600,000 through gift cards, a court heard.
A gang of four was sentenced on Tuesday for the plot, where ‘SMS blasters’ hidden inside suitcases were wheeled around the Underground network and sent out phishing texts.
Travellers who walked past the devices received fake messages about a failed parcel delivery and a link inviting them to enter their details to sort the problem.
xxx
xxx
But a growing number of forecasters now say the economy’s dependence on AI was overstated. Prominent economists, including from Morgan Stanley and JPMorgan Chase, calculate that the AI buildup was directly responsible not for 92 percent or 39 percent of gains to the U.S. economy in 2025, but as little as zero.
From: How much did AI boost the economy? Maybe zilch, some economists say. – The Washington Post.
xxx
xxx
Other than being incredibly entertaining and slightly worrying for those concerned about AI gaining sentience, OpenClaw and Moltbook offer a glimpse of where AI is going.
Actually, that wasn’t the lesson I took away from my first look at what was going on over there. The lesson that I took away from (you will not be surprised to hear) is that without a working digital identity infrastructure, we can’t have nice things.
xxx
There are 1.5 million agents transacting on Moltbook right now. Depending on who you ask, this is either the early singularity, a dumpster fire, or 17,000 humans puppeting bots. Wiz Research found 341 malicious skills on ClawHub stealing credentials. Karpathy called it what it is.
From: What Agents Need Before They Handle Real Money – Catena Labs.
xxx
xxx
Even if Moltbook does not spell the imminent subjugation of humanity, it poses other risks. Some careless users are running up thousands of dollars in cloud-computing fees as their agents draw on cutting-edge ai models to function. Then there are the scammers, who are taking advantage of the free rein that OpenClaw agents have over the devices on which they run on. Already Moltbook has been inundated by attempts (including by humans pretending to be bots) to convince ai agents to hand over cryptocurrency. The strange experiment could well prove costly—and short-lived
From: A social network for AI agents is full of introspection—and threats.
xxx
\
xxx
merchants increasingly treat unidentified automation as a policy and risk problem. Amazon’s lawsuit against Perplexity over agentic shopping makes the direction clear.
– The path forward is not “smarter bots clicking websites.” It’s agent-native commerce interfaces: OpenAI and Stripe’s ACP, Google’s UCP, and browser-level standards like WebMCP.
xxx
xxx
merchants increasingly treat unidentified automation as a policy and risk problem. Amazon’s lawsuit against Perplexity over agentic shopping makes the direction clear.
– The path forward is not “smarter bots clicking websites.” It’s agent-native commerce interfaces: OpenAI and Stripe’s ACP, Google’s UCP, and browser-level standards like WebMCP.
xxx
xxx
There are 1.5 million agents transacting on Moltbook right now. Depending on who you ask, this is either the early singularity, a dumpster fire, or 17,000 humans puppeting bots. Wiz Research found 341 malicious skills on ClawHub stealing credentials. Karpathy called it what it is.
From: What Agents Need Before They Handle Real Money – Catena Labs.
xxx
Snippings