> daily_signal(2026_11_04)
Huawei is offloading the sanctions risk of its own AI chips onto customers, as a nudify tool hides in a vetted app, a coding agent refuses to clone games, and an AI takes a telescope's controls.
PickBits Daily Signal · Wednesday, November 4, 2026
// tl;dr
- Huawei is telling some of its own Ascend customers to source the high-bandwidth memory for its AI chips themselves. It can't get enough domestically and U.S. export rules block it from buying from Samsung and SK Hynix, who together make over 80% of HBM. That pushes buyers onto a gray market where one Shenzhen broker puts a smuggled HBM3E stack at about $4,000, up from a few hundred dollars a year ago, and at six to eight stacks per chip the memory alone can reach roughly $32,000.
- A Wired investigation found hundreds of ordinary-looking App Store and Play apps hiding AI "nudify" deepfake tools, built to stay invisible to the reviewers who approve them. The trick: the app's real interface simply doesn't load for anyone inside the app-store reviewer's region. We covered a smaller version when the Tech Transparency Project found 31 nudify apps rated for minors and Apple pulled 28; Google suspended several too.
- Anthropic added two sentences to Claude Code's Opus 5.5 system prompt that make it refuse to recreate existing commercial games. Anthropic's Addy Osmani confirmed the change, which followed a wave of users mass-producing unofficial browser ports of games like Halo and GTA Vice City. It applies even to a clone renamed as an "original," and some users report it falsely blocking legitimate original work.
- SETI and Breakthrough Listen handed an agentic AI the Allen Telescope Array's real-time detection pipeline, running on NVIDIA's Holoscan, while researchers still decide where to point the dishes. It caught the Crab Nebula pulsar's giant radio pulses as the data streamed in, doing in hours what used to take weeks. No extraterrestrial signal has turned up yet.
Export controls were meant to stall China's AI chips. We've watched China route around them all year, from the chip-workaround reporting we covered in the spring to the giant Huawei cluster reported behind DeepSeek. This week it's the memory. Huawei can build the logic now; what it can't get is the high-bandwidth memory bolted to it, so it's telling its own customers to go buy that memory on a gray market it's legally barred from touching. The shortage didn't ease. It just moved to the one part you can only get by smuggling it in.
The nudify one made me angry: Wired found hundreds of ordinary-looking apps hiding those tools and built to slip past Apple and Google's review, the same rot we covered when 31 of them turned up rated for minors. Anthropic went and wired Claude Code to refuse to clone games, then promptly tripped its own rule and blocked real projects. And the one that made me smile was out of California, where researchers handed an AI a radio telescope and it caught a pulsar in real time, the kind that used to take weeks to find.
Huawei's own buyers are now the ones smuggling in the memory its chips need, at about four thousand dollars a stack.
1. Huawei is telling its own AI-chip customers to buy the memory themselves.
Barred from Samsung and SK Hynix by U.S. export rules, it is pushing some Ascend buyers onto a gray market where a smuggled HBM stack now runs about $4,000.
Huawei is now telling some Ascend AI-chip customers to source their own high-bandwidth memory, because it can't secure enough at home and can't legally buy it abroad. The reason is the memory, not the logic die. Samsung and SK Hynix together control more than 80% of the world's HBM, and U.S. export rules bar Huawei from buying from either. So the memory flows through a gray market instead. One Shenzhen broker puts a smuggled HBM3E stack at about $4,000, up from a few hundred dollars a year ago, and at six to eight stacks per chip the memory alone can reach roughly $32,000 per accelerator.
For a buyer, this quietly makes you the importer of the one component customs authorities are watching. South Korea's customs service had already flagged $560 million in trade violations by May, a third of it in strategic goods. The company that sold you the chip just handed you its sanctions problem. "China is catching up on AI chips" has been the headline for two years, but China can fabricate the processor and still can't make, at volume, the memory that processor is useless without.
Why this matters: When I have asked a vendor where a component actually comes from, the honest answer is rarely in the contract until you force it there, and nobody writes down a gray-market HBM source on their own. On these chips the memory, not the processor, is now the hard part to get, and it is the piece most likely to land you in a sanctions mess. The "catching up" story hides the real tell: if you're making your buyers go find the parts themselves, you can't get them either.
Action this week: Treat the memory, not just the accelerator, as the part that can land you in a sanctions mess on any China-sourced hardware. Ask where the HBM in any Ascend-based system comes from, and get the answer in writing before you sign, because "sourced by the customer" is a liability you would be signing for. Go to the rules directly: the U.S. BIS Entity List and HBM rule at bis.gov and TrendForce's HBM coverage at trendforce.com show which parts and prices are moving gray-market.
the-decoder.com: Huawei tells customers to find their own AI-chip memory, effectively outsourcing procurement to the black market (November 3, 2026)
tomshardware.com: Huawei backs development of HBM memory in China; new consortium aims to sidestep U.S. sanctions (2026)
the-decoder.com: Huawei's AI-chip production boom reportedly faces a critical shortage of high-bandwidth memory (2026)
2. Hundreds of ordinary-looking apps are hiding AI nudify tools from the reviewers who approved them.
A Wired investigation found photo editors and utilities concealing deepfake "nudify" features, with one trick being an interface that won't render inside the app reviewer's own region.
The whole point of app-store review is to catch this, and these apps are engineered specifically to get past it. Hundreds of innocuous-looking apps in the Apple and Google stores conceal an AI "nudify" capability that generates sexualized images of real people, and the abusive features are deliberately hidden from the reviewers who approve them. The trick is brazenly simple: the app's real interface doesn't load for anyone sitting in the reviewer's region, so the person approving it never sees the tool they're approving.
This isn't new ground for us. The Tech Transparency Project earlier found 31 nudify apps rated as suitable for minors; Apple said it removed 28 of the apps named in that report and Google said it suspended several. This isn't a handful of apps that slipped through. It's hundreds, built from the start to beat the one gate standing between a nudify tool and a phone a parent handed to a kid. This isn't buried on some dark-web forum that takes effort to find. It's one tap deep inside an app that cleared Apple's own review.
Why this matters: The app that passed Apple's review may be hiding a nudify tool on the phone in your kid's hand, and the store's green light is exactly the trust these apps hide behind. A review an app can switch off by geofencing the reviewer isn't protecting anyone, and it's kids and anyone with a photo on a public profile who get hurt. What gets me is that "it's in the App Store" has been most families' entire safety check, and it just failed on purpose.
Action this week: Report a nudify or deepfake app directly in the store, through Apple's Report a Problem or Google Play's "Flag as inappropriate," and lock down installs on any device a minor uses with Screen Time on iOS or Family Link on Android. If a real person has been targeted, take the image down through StopNCII, or NCMEC's Take It Down for anyone under 18, and file with the CyberTipline. Send this to whoever sets the device rules in your house.
wired.com: Hundreds of mundane AI apps are hiding abusive deepfake features (November 2026)
engadget.com: Apple and Google reportedly still offer dozens of AI nudify apps (2026)
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
3. Claude Code now refuses to recreate existing games.
Anthropic added two sentences to the Opus 5.5 system prompt after users mass-produced AI-built knockoffs of Halo and GTA Vice City, and the block applies even to a clone renamed an "original."
If you opened Claude Code this week to rebuild a favorite game, it said no. Anthropic added two sentences to Claude Code's Opus 5.5 system prompt that make the tool refuse to recreate existing commercial games, and employee Addy Osmani confirmed the change: using Claude to infringe intellectual property violates the usage policy. The trigger was a wave of users spinning up hundreds of unofficial browser ports and clones of titles like Halo and GTA Vice City, plus knockoffs of apps such as Adobe's. The rule had been live for several days by the time it was reported, and it holds even when a clone is renamed and passed off as an "original."
The same guardrail is already catching legitimate, original work. Some users report false positives blocking projects that infringe nothing, which is going to happen when you leave it to a model to eyeball what counts as a clone. Originals built from your own assets still go through. A clone with a new name on it doesn't, and the rename isn't a workaround, it's the exact case Anthropic says it wrote the rule to stop.
Why this matters: My own read: the precedent matters more than the games. Your coding agent's vendor just changed what the agent will build, mid-stream, without asking you, and it can do that again tomorrow on something you depend on. "The model generated it" has never been an IP shield, and now Anthropic has baked that into the tool at the system-prompt level. The false-positive reports worry me more than the games do: a guardrail strict enough to stop a clone can stall a real project on a bad guess.
Action this week: Re-read your AI vendor's acceptable-use policy, Anthropic's AUP included, and make it plain internally that "the model wrote it" doesn't clear IP review. The part I'd actually plan for is the false positive: give an engineer a fast way to flag a legit project the agent wrongly blocks, because that is the failure that will eat a day. And Anthropic says it'll keep tuning the rule, so this won't be the last time the agent's behavior shifts under you. If you own AI tooling on your team, forward this to them.
the-decoder.com: Anthropic cracks down on game cloning in Claude Code after wave of AI-built knockoffs (November 3, 2026)
anthropic.com: Usage Policy (Acceptable Use Policy)
4. An AI is now steering SETI's hunt for alien life.
SETI and Breakthrough Listen put an agentic AI on NVIDIA's Holoscan in charge of the Allen Telescope Array, and it caught the Crab Nebula pulsar's giant pulses as the data streamed in.
This is the kind of thing I actually want automation doing. The SETI Institute and Breakthrough Listen have put an agentic AI in the loop at the Allen Telescope Array: it steers the dishes, corrects the data, and runs detection models in real time on NVIDIA's Holoscan platform, with the hardware able to handle up to 2.4 terabits per second. The system caught the Crab Nebula pulsar's giant radio pulses as the data streamed in, rather than in weeks of offline analysis after the fact.
No alien signal turned up. The target was the Crab Nebula pulsar, something astronomers already know cold. People still enter the commands; the AI runs the pipeline, it doesn't pick where to look. Weeks of radio-astronomy triage now take hours, and the same system that catches a pulsar mid-stream helps pull any faint signal out of the noise.
Why this matters: What gets me here is the time it buys back. No signal yet, and a person is still in the chair, but work that ate weeks now takes hours. That is the kind of AI I want more of. It keeps pace with more data than any team could realistically read, and I trust it more because the researchers told me what it can't do yet, not just what it can.
Action this week: Follow the open-science trail. The SETI Institute and Breakthrough Listen publish their data and tooling, and the team says it's evaluating open models it can run locally at the edge. Watch whether that real-time pipeline gets released or stays in-house, because an open version is the one that lets a university, or a curious amateur, point the same technique at their own noisy data.
the-decoder.com: AI agents take over at SETI's Allen Telescope Array, scanning for extraterrestrial intelligence (November 3, 2026)
developer.nvidia.com: How SETI uses AI to search for intelligent alien life (2026)
breakthroughinitiatives.org: Breakthrough Listen and real-time detection at the Allen Telescope Array (2026)
» What to watch this week
- Watch whether a domestic HBM supply actually ships for Huawei, or "source it yourself" just becomes how it works from now on. There's a Chinese HBM consortium meant to sidestep sanctions; the tell is whether any of it reaches volume, or the gray market simply gets bigger.
- Apple and Google's real test isn't which apps they pull after a report. It's whether either store closes the geofenced-reviewer hole that let these tools pass review in the first place, because removing 28 apps after the fact doesn't touch the technique.
- Watch how many legitimate projects Anthropic's game-cloning rule snags. Anthropic says it'll keep tuning the block, so see whether the false-positive reports grow or settle, and whether other coding-agent vendors copy the move.
- SETI's pipeline is the one I'd keep an eye on. The team says it's evaluating open models it can run at the edge, and an open release is what turns one California array's breakthrough into a tool other observatories, and other sciences, can pick up.
Tomorrow's signal lands here.