Skip to content
LiveNext 14:18:33
Today's five/N°47 · Sunday, November 22, 2026/Filed 03:49 ET

AI news today: AI providers leaked prompt secrets, Lean certifies the Riemann proof

One story asks you to rotate your keys today. The other four will take months or years to settle.

The five · ranked by consequence

Permalink ↗
  1. Together AI confirms prompt leak after Dragos test

    Together AI says an intruder downloaded stored prompt samples from June and July, and OpenRouter has cut three providers after Dragos seeded nearly two million prompts with fake secrets. If a key ever went into a prompt, assume someone read it.

    Why it matters If a credential has ever gone into a prompt, a pasted log or an agent's context, rotate it this week. Then work out which companies in your AI stack actually see your requests, because the privacy setting you ticked only protects you at the first one.

    SRCDragos

  2. Lean community certifies Claude-aided Riemann proof

    The Lean and Mathlib communities say they have "full confidence" that Levent Alpöge's 17-million-line Lean proof of the Riemann Hypothesis is correct, after several independent checkers accepted it. The machine half of the question is answered. The human half has barely started.

    Why it matters Yesterday's open question was whether Lean proved exactly what Riemann stated. The people who maintain Lean now say it did. The lesson for anyone using AI is the same one, sharper: this output earned trust only because several checkers built by different people all agreed.

    SRCLean FRO and Mathlib Initiative

  3. UN talks fail to launch an AI weapons treaty

    After all-night talks in Geneva, the 128 states in the Convention on Certain Conventional Weapons agreed on Nov. 21 to "elaborate" an instrument on lethal autonomous weapons over two years, with no commitment that it will bind anyone. The word they dropped was "negotiate," and that is the outcome.

    Why it matters No treaty will limit AI in targeting for at least two more years. Until then the working limits on military uses of AI are national rules and the usage policies of the companies that build the models, so if you build on a frontier model, its policy is the rulebook you actually live under.

    SRCICRC

  4. Korea's education minister quits over AI exam cheating

    South Korean Education Minister Choi Kyo-jin resigned on Nov. 22 over rushed measures against AI-assisted cheating on the Suneung college entrance exam, where 111 test takers were caught with smart devices. Every school and certification body now has his problem.

    Why it matters If you run a test, a certification or a hiring screen that assumes the person answering is alone, Korea just showed what it costs to fix that the week of the exam. Redesign the assessment now, before your own scandal does it for you.

    SRCThe Korea Herald

  5. Tesla starts FSD v15 rollout with live Reality View

    Tesla began rolling out FSD (Supervised) v15 to Early Access members on Nov. 21, claiming "25x better handling of safety critical cases" and adding a live view that Tesla says lets you see and control what FSD plans to do. The view is the real news. The 25x is Tesla's number until someone else measures it.

    Why it matters Reality View shows the driver what the system plans before it acts and lets them change it. That is the same see-the-plan, steer-the-plan pattern every AI agent needs. Wide release is promised for December, so judge v15 on independent intervention data, not a launch post.

    SRCTeslarati

Prompt of the day

Find the secrets hiding in your AI workflow

› You are a careful security reviewer. I want to find every place a secret could leak through the AI tools I use.

Here is how I use AI today:
[List your tools: chat apps, coding agents, API providers, routers, automations, and what you paste into them]

Do three things.

1. Make a table of every point where a credential, API key, password, token, customer record or private URL could end up inside a prompt, a pasted log, a file I upload, or an agent's context. For each row, name the tool, what could leak, and how likely it is in my setup.

2. For each tool, list which company actually receives the data, including any provider a router or marketplace forwards it to. Tell me which of these I should check for a zero data retention or no-training setting, and what to look for in their settings or docs.

3. Give me a short, ordered checklist for this week: which keys to rotate first, how to keep secrets out of prompts (environment variables, a secrets manager, scoped short-lived tokens), and how to scan anything an AI coding agent commits before it is pushed.

Ask me up to three questions first if my description is too thin to do this well. Be specific to my setup, not generic.

Works in any AI chat. In Magai, run it against several models at once.

Every day, on the record.

Each square is an edition. Click any day to read the five that mattered then.

Editions
47
Stories ranked
235
Topics tracked
10
Terms defined
220
MayJulAugSepOctNov
EditionNo editionTodayFull archive →

Speak fluent AI.

Every term our stories use, defined in plain English and linked wherever it appears.

The glossary →