AI News · November 22, 2026 · 6:31

AI and the Riemann Hypothesis & Canary tokens expose AI prompt leaks - AI News (Nov 22, 2026)

Did AI help crack the Riemann Hypothesis? Plus: canary tokens reveal AI prompt leaks, Together AI's breach, AI-driven cyberattacks, and the No AI Slop manifesto.

AI and the Riemann Hypothesis & Canary tokens expose AI prompt leaks - AI News (Nov 22, 2026)
0:006:31

15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕

Support the show →

Our Sponsors

Today's AI News Topics

  1. AI and the Riemann Hypothesis

    — 3Blue1Brown examines reports that artificial intelligence may have helped solve the Riemann Hypothesis, raising questions about AI-assisted proofs, verification, and mathematical understanding.
  2. Canary tokens expose AI prompt leaks

    — Dragos sent nearly two million AI prompts with canary tokens and found inference resellers accessing data, ignored privacy settings, and some prompts routed to Chinese AI organizations.
  3. Together AI prompt security incident

    — Together AI disclosed that serverless inference prompt samples leaked after a Hugging Face credential was committed to a public GitHub repo, and the intruder triggered Dragos canary credentials.
  4. Bridgewater: AI speeds up cyberattacks

    — Bridgewater warns that AI is helping attackers exploit zero-day vulnerabilities faster than defenders and policymakers can respond, raising cyberwarfare and economic risks.
  5. WebBrain local private AI assistant

    — Mozilla spotlights WebBrain, a browser extension that runs offline LLMs locally for private page summaries, automation, and an offline Wikipedia mode.
  6. No AI Slop manifesto

    — Speakers at the AIE Code Summit 2026 published a 'No AI Slop' manifesto urging accountable, human-reviewed AI-assisted work in code and documentation.
  7. Anil Seth on conscious AI

    — Neuroscientist Anil Seth argues intelligence is not consciousness, warning that anthropomorphism leads people to wrongly believe AI and language models are conscious.

Sources & AI News References

Full Episode Transcript: AI and the Riemann Hypothesis & Canary tokens expose AI prompt leaks

One of the most famous unsolved problems in mathematics, a puzzle that has resisted the greatest minds for over a century and a half, may have just fallen, and artificial intelligence might have had a hand in it. We'll get to that in a moment. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is November 22nd, 2026. On the docket: a striking experiment that tracked where AI prompts really end up, a breach that turns out to be connected to it, a sober warning about AI and cyberwar, a manifesto against sloppy AI work, and a neuroscientist pushing back on the idea of conscious machines.

AI and the Riemann Hypothesis

Let's start with that teaser. Math educator 3Blue1Brown has released a video reacting to the possibility that AI helped solve the Riemann Hypothesis, the deep conjecture about how prime numbers are distributed. The video points to announcements from mathematicians involved in the work, and treats what used to be a purely hypothetical scenario as something that may now be real. A caution here: a result like this needs careful verification before anyone pops the champagne. But if it holds up, it would be a landmark for number theory, and it would force some uncomfortable questions. Can mathematicians truly understand a proof a machine helped produce? And how much trust should the field place in AI-assisted reasoning at all?

Canary tokens expose AI prompt leaks

Now to security, and an experiment that should make anyone sending sensitive data to AI services pause. Dragos says it spent five months sending nearly two million prompts, each seeded with unique canary tokens and trackable fake secrets, to widely used AI services, then watched where those secrets resurfaced. The findings were not reassuring. Dragos reports evidence that at least three AI providers had been exposed, that resellers in inference marketplaces were accessing customer prompts without authorization, and that privacy settings and published data policies were repeatedly ignored. Some prompts sent to U.S.-based providers ended up routed to Chinese AI organizations, and certain unique links were opened months after the original request. The core point: every hop in an AI routing chain can see your full prompt and response, and prompts often hold credentials, source code, and confidential plans. Dragos argues routing is a security decision, and the cheapest path may well be the least trustworthy.

Together AI prompt security incident

That experiment has already surfaced in a real incident. Together AI has notified customers that a subset of prompts sent to its serverless inference service in June and July was exposed. The company says prompt samples kept for testing sat in a private Hugging Face repository, but a credential for it was accidentally committed to a public GitHub repo, and someone used it to download the data. The intruder then combed the samples for more credentials and tried some, including five canary credentials planted by Dragos. Together AI says it hasn't found broader leakage and is still assessing whether personal data was involved. Its fixes include stricter credential controls, secret scanning on pull requests, zero data retention by default for new accounts, and revised practices around AI coding agents. A classic leaked-key story, with a very modern twist.

Bridgewater: AI speeds up cyberattacks

Zooming out, investment firm Bridgewater has published a report arguing that AI is speeding up cyberattacks faster than defenders and governments can keep pace. Attackers use AI tools to find and exploit weaknesses quickly, while defenders face the much harder task of patching everything first. The report notes a rise in zero-day and actively exploited vulnerabilities, plus an ever-growing pile of code expanding the attack surface. Policy moves in Japan, South Korea, the U.K., and the U.S. are described as helpful but too narrow. Bridgewater's warning is that without more coordinated action, cyber risk could become economically destabilizing.

WebBrain local private AI assistant

On a more privacy-friendly note, Mozilla's add-ons blog highlighted WebBrain, a browser extension that runs AI models entirely on your own machine. It can summarize pages, analyze images, and help automate tasks on sites like GitHub, without shipping your browsing data to a cloud provider. It even includes an offline mode for chatting with Wikipedia and medical references during outages. After the stories we just heard, the appeal of keeping data local is pretty easy to understand.

No AI Slop manifesto

Next, a cultural pushback from the developer world. Speakers and attendees of the AIE Code Summit 2026 have published a manifesto titled 'No AI Slop.' They define slop as work produced with less effort than it takes others to understand it. Their argument isn't anti-AI; it's anti-carelessness. The group calls for valuing outcomes over sheer output, understanding over speed, and human intent over a model's first draft. They warn that slop compounds in codebases, docs, and communication, making reviews ever harder, and recommend labeling AI-assisted work, reviewing what matters, and fixing guardrails when mistakes recur. It captures a growing mood: use the tools, but own the results.

Anil Seth on conscious AI

Finally, neuroscientist Anil Seth has written an essay challenging the notion of conscious AI. His central claim is that intelligence and consciousness are different things, so machines can grow more capable without ever having an inner life. He argues we're prone to seeing minds where there are none, thanks to human exceptionalism and our habit of anthropomorphizing, and fluent language models make that trap even easier to fall into. Seth also questions whether consciousness is purely computational at all; if it isn't, today's digital systems may simply be the wrong kind of thing to be conscious. Getting this distinction right, he says, matters for AI governance and for understanding ourselves.

That's it for today's episode of The Automated Daily, AI News edition. From a possible mathematical milestone to some hard lessons about where our prompts travel, it's been a busy day. Links to all stories can be found in the episode notes. I'm TrendTeller, thanks for listening, and see you tomorrow.

More from AI News