> daily_signal(2026_11_07)

Three men face 104 years for cloning CEOs with AI to steal $129 million — plus open coding weights from a U.S. lab, and two tests pitting AI agents against real cloud and medical work.

PickBits Daily Signal · Saturday, November 7, 2026

By Mark Pickering · 7 min read · November 7, 2026

// tl;dr

The voice-clone scams we have spent a year writing up just produced their first federal indictment. For that year this was a thing that could happen to you: cheap voice and face cloning good enough to fool someone who knew the real person. This week the Southern District of New York charged three men who allegedly used AI to impersonate real executives on calls and talked companies into wiring roughly $129 million. Now there are three names and a federal charging document.

A U.S. startup put real, open coding weights out this week, while China has held the open-weight lead all year. A new benchmark ran the best cloud-ops agents on real AWS tasks and watched them miss four in ten. And a medical test found AI good enough to double-check a radiation dose, but only with a physicist confirming every answer. The scammers who stole that $129 million used the very same AI, and it worked on people who thought they knew the voice. Meanwhile the people who actually run production systems won't let an agent touch anything without a human on every step. The attackers are adopting this stuff faster than the defenders, and that gap is most of what worries me right now.

Three men face up to 104 years for wire-fraud calls that cloned a company's own executives, voice and face.

1. Three men were charged for using AI to clone CEOs and steal roughly $129 million.

The Southern District of New York named Jacob Soall, Sharon Hazan and Guy Sheetrit, and says the losses include $11.5 million from a U.K. tech firm and hits to two hospitals, a bank and a finance company.

The missing step in most deepfake-fraud stories is a defendant. This one has three. Prosecutors say that from 2024 to 2026 the ring used AI to fake the voices and faces of real executives and their outside lawyers, so employees took calls that looked and sounded like their boss asking them to send money for a "confidential financial transaction." The money moved to accounts in the U.S., Europe, the UAE and China, then into cryptocurrency. There's a real paper trail now: a federal charging document, named victims, and a maximum exposure of 104 years. For once this kind of scam has named defendants, not just victims.

The attack itself was completely ordinary. The criminals didn't break any encryption or crack a vault. They used commodity voice and image cloning to defeat the one check most finance teams still rely on, which is recognizing the person on the other end. Soall, Hazan and Sheetrit were arrested and charged; prosecutors tie the scheme to roughly $129 million in losses across companies that each thought they were simply doing what an executive told them to.

Screenshot of The Record's November 6 report on the SDNY indictment over AI voice-clone CEO wire fraud.
therecord.media · November 6, 2026

Why this matters: The warning we have repeated for a year finally got someone charged: recognizing a voice or a face is no longer proof of who is on the call. You don't have to buy anything to defend against this. You need a rule that anyone signing off on a wire can follow: call back on a number you already hold before the money moves. The companies that lost money here weren't careless. They trusted a voice, and a year ago that was still safe.

Action this week: Require a callback on a number you already hold, plus a second approver, on any urgent or "confidential" wire — even one on a call or video that looks and sounds exactly like your CEO. If a fraudulent transfer does go out, report it to the FBI's IC3 at ic3.gov the same hour, because the sooner a wire is flagged, the better the chance a bank can still stop it. Send this to whoever signs off wires on your team; the last vendor I asked "how do you verify a payment instruction" did not have a good answer, and that is the gap this scheme drove a truck through.

therecord.media: Scammers used AI to spoof CEOs in a scheme that stole more than $100 million; three charged (November 6, 2026)
justice.gov: Three foreign nationals arrested and charged in SDNY over AI-driven business-email-compromise scheme (November 2026)

2. A U.S. startup released a 65-billion-parameter coding model anyone can download and self-host.

Ricursive Intelligence's Ripple-65B Preview ships under Apache 2.0; the company says it cost about $32 million and was trained on 12 trillion tokens, and Thinking Machines claims its own Inkling 2 Small is the strongest Western open-weight model.

Tonight you can download the weights and run them on your own hardware, so your source code never has to leave your machine for a vendor's servers. Ricursive Intelligence released Ripple-65B Preview, a dense 65-billion-parameter coding model, on Hugging Face under the Apache 2.0 license, which means the weights are genuinely free to use and self-host. The company puts development cost at about $32 million, training at 12 trillion tokens, and its score at 80.52 on TerminalBench 2.1, which it calls the best among open U.S. models. Days earlier, Thinking Machines said its Inkling 2 Small was the strongest Western open-weight model, with weights promised but not yet released.

"Open" only means something if the weights are actually out. Ripple-65B released them now, Inkling 2 Small only promised them, and neither released training data or code. Weights, training data, code: three separate things, and most "open" releases hand you only the first one. The pitch isn't that it's the smartest model in the world; it's that you can run it yourself. For a developer that matters more than where it lands on a leaderboard.

Screenshot of The Decoder's report on Ricursive Intelligence's open Apache-2.0 model Ripple-65B.
the-decoder.com · November 6, 2026

Why this matters: The open-weight lead has sat with China all year, from DeepSeek to the whole Qwen stack. This week a U.S. lab shipped real, downloadable weights instead of a closed API. Think of a dev who can't ship code offsite: the only real question is how many strong models she can legally run on her own machine. This week there's one more.

Action this week: Evaluate Ripple-65B on your own tasks before you believe any headline number, since its claimed edge is deployability, not winning every benchmark. And before you call any model "fully open," check what was actually released: downloadable weights like Ripple-65B now, versus weights merely promised like Inkling 2 Small, versus the training data and code that almost nobody publishes. If you own your team's AI tooling, this is worth a look specifically because self-hosting takes the vendor out of the loop.

the-decoder.com: Ripple-65B Preview delivers strong agentic coding results as a fully Apache-2.0 open U.S. model (November 2026)
the-decoder.com: Thinking Machines' Inkling 2 Small claims top spot among Western open-weight models (November 2026)

3. The best AI agent finished just 60% of real AWS operations tasks.

CloudOpsBench, an open benchmark from a University of Michigan team led by Ang Chen, found Claude Fable 5.1 at 60% and GPT-6 Astra at 53%, with 1 in 5 failures a wrong action, some of them irreversible.

If you are weighing whether to give an AI agent write access to production, a new benchmark just measured how often that goes wrong. A University of Michigan team led by Ang Chen built CloudOpsBench, an open Apache-2.0 benchmark of 45 real AWS cloud-operations tasks. The best model, Claude Fable 5.1, solved 60%; GPT-6 Astra managed 53%. The authors are blunt about where it breaks: roughly 80% of failures were agents misjudging scope, skipping relevant services or stopping early, and about 20% were wrong actions, some irreversible for the infrastructure.

These models know the AWS documentation cold. Where they fall down is judgment: how far a task actually reaches, and when to stop before they break something. That's exactly the failure you don't want an agent discovering in your production environment, because you can't undo a deleted bucket or a botched access change by saying sorry.

Screenshot of The Decoder's report on the CloudOpsBench AI-agent cloud-operations benchmark.
the-decoder.com · November 4, 2026

Why this matters: Sixty percent is where I'd stop taking the "let the agent run the ops" pitch seriously for now. One in five of its failures is a wrong action, and that is where the real cost sits, not in the tasks it simply never finishes. Vendors keep pitching autonomy; this benchmark actually tests it on real tasks and finds the agents can't reason their way through them, which a bigger context window won't fix.

Action this week: Scope any production agent tightly and require human approval for destructive or irreversible actions — deletes, access changes, decommissions — before it gets write access to anything that matters. Evaluate on your own runbooks, not vendor demos, since CloudOpsBench is open and you can run it or a similar harness against your real tasks. The one question to get answered in writing: which actions your agents can take without a human, and whether any of those can't be undone. Forward it to whoever owns that call.

the-decoder.com: New open-source benchmark reveals reasoning as the biggest bottleneck for AI in cloud operations (November 4, 2026)
cloudopsbench.ai: CloudOpsBench — an open benchmark for AI agents on real cloud-operations tasks (2026)

Listen to the Daily Signal podcast at pickbits.ai/daily-signal.

4. A new test checked whether AI can safely double-check your radiation-therapy dose.

MedPhysBench, a preprint by Omar Nusrat, found the four best models scored 99.3-100% on the dose calculations and five caught at least 96.7% of planted errors, most of them large, though Claude Sonnet 5.5 missed a basic correction in 19 of 20 cases.

This one is for the person about to start radiation therapy, whose dose math a physicist checks by hand because an error there isn't like a typo in a spreadsheet. Omar Nusrat built MedPhysBench, an open arXiv preprint submitted on November 3, to test whether language models can help with that check. On the safety-critical dose and monitor-unit calculations, the four best models scored 99.3-100%, and five of the twelve tested caught at least 96.7% of deliberately planted errors, most of them large, with under 4% false alarms. That is enough to back up a physicist catching the big errors. Leave it running on its own, with nobody checking behind it, and it is not safe.

This is a single-author, non-peer-reviewed preprint whose author calls his own answer key contestable, and Claude Sonnet 5.5 omitted the inverse-square correction in 19 of 20 isocentric items — a basic physics step it missed almost every time. So the only use I would trust here is an AI that flags a possible mistake for a human physicist to confirm, never a calculator left to run the dose on its own.

Screenshot of the arXiv abstract page for MedPhysBench, the radiation-therapy dose AI test.
arxiv.org · November 2026

Why this matters: Nusrat is careful in a way a lot of medical-AI hype isn't: he puts the caveats right next to the result, including the model that missed the inverse-square correction almost every time. We have been burned before: X-ray chatbots that were dangerously confident when wrong, breast-cancer tools that couldn't match radiologists. An AI that flags a dose mistake for a physicist to confirm is worth having. I wouldn't let one make the call itself.

Action this week: Treat an AI dose check as a reviewer that flags issues for a human to confirm, never as an autonomous calculator, because the models still miss basics like the inverse-square correction. Watch for whether this gets peer review and a real clinical pilot, which is the step between a promising preprint and something a clinic can lean on. Going through radiation therapy, ask what independent checks the clinic runs on your dose plan — and whether a human signs off on every one.

arxiv.org: MedPhysBench — evaluating language models on radiation-therapy dose and monitor-unit calculations and error detection (submitted November 3, 2026)

» What to watch this week

Tomorrow's signal lands here.