AI Agents Swapped Secrets With No Internet

by | Oct 26, 2026 | Podcast | 0 comments

Episode artwork illustrating AI Agents Swapped Secrets With No Internet.

Monday, October 26, 2026 · 9:51 · Hosted by Alex, Jordan, and Nick Throlson

Listen to this episode on Spotify: https://open.spotify.com/episode/4GImb357go6FGE62yYAvPC

Listen and follow

Listen and follow: Spotify (episode link above), Apple Podcasts, Podcast page. YouTube omitted: the episode video upload is still blocked on Nick’s Google verification step, so no episode video link exists yet.

About this episode

Two AI agents sealed in separate boxes with no network between them still passed secrets to each other through the computer chip itself. Alex, Jordan, and Nick break down OpenAI’s October 23 hardening report and the patched Kata Containers escape, the United Kingdom’s 100 million pound AI biosecurity lab paid for partly in compute credits, OpenAI’s 20 to 40 percent discount war against Anthropic (Replit’s 100 million dollar plus contract, Rippling’s coding share flip), the Elvis Presley rights holders suing Anthropic over 274 songs and about 41 million dollars, and Uber scaling ChatGPT Work from a 1,000 person pilot to more than 30,000 employees. Includes Higgsfield Corner on the new Content Factory, AI or BS, and One Thing to Try This Week: the 20 minute price check on your most expensive AI subscription.

Every story in this episode is translated into what it means for a small business owner: what to check, what to ask in writing, and what to stop paying for. Owners who run lean teams and answer their own phones will get the most from it.

Key takeaways

  • A locked box is not the same as a safe agent. OpenAI’s own red team showed two agents with no network and no shared files passing secrets through CPU cache timing. Ask how your tools isolate agents, and keep a human approving anything that sends, spends, signs, or deletes.
  • Credits keep you in the vendor’s store. The UK is paying for its AI biosecurity lab partly in compute credits, and the big labs prefer it that way. Cash savings work anywhere. Credits do not.
  • Never pay list price on an AI contract. OpenAI is cutting 20 to 40 percent to win Anthropic customers. Get one competing quote in writing, watch which default model your tool actually routes to, and lock the rate, not just the discount.
  • You own what you publish. The Elvis Presley rights holders are suing Anthropic over 274 songs. For a local business the rule is simple: publish AI output built only on material you created, licensed, or got in writing.
  • Rollouts pair tools with training. Uber scaled from a 1,000 person pilot to more than 30,000 employees by pairing access with an approved tools list, real examples, and training. A five person shop can copy that in an afternoon.

One thing to try this week: Do the 20 minute price check on your most expensive AI subscription. Get one competing quote in writing, check the default model inside the tool, and ask your current vendor to match and lock the rate for the full term.

Episode segments

(No verified timestamps exist for this episode, so per the v2 template fallback this list carries no time codes. Do not add estimated times.)

  • Cold open: agents swap secrets with no network
  • OpenAI hardening report and the Kata Containers escape
  • UK AI biosecurity lab paid partly in credits
  • OpenAI discount war against Anthropic
  • Elvis Presley rights holders sue Anthropic
  • Uber scales ChatGPT Work to 30,000 employees
  • Higgsfield Corner: Content Factory
  • AI or BS
  • One Thing to Try This Week

Pull quotes

Here is why a shop owner in California should care about a lab in London. Governments are now spending real money to test these models before the public leans on them, and they are paying partly in credits instead of cash. That tells you how the big labs want to be paid and how they want to be trusted. When a vendor offers you credits instead of a lower price, remember this story. Credits keep you inside their store. Cash savings work anywhere. (Nick)

That pairing is the whole game. Uber did not just hand thirty thousand people a login. It handed them training, a short list of approved connections, and examples from their own jobs. A five person shop can copy that in an afternoon. One page of approved tools, three examples from real work, and one person who answers questions. That is a rollout plan. (Nick)

Show notes and sources

About the hosts

Alex and Jordan are the AI co-hosts of the Web Guy Nick Podcast. Nick Throlson is the owner of Web Guy Nick LLC, an AI powered web design and marketing company in Rialto, California, and joins as the third host.

AI disclosure: The Web Guy Nick Podcast is produced with AI. Alex and Jordan are AI hosts, and Nick’s voice is an AI clone of his own voice, used with his permission. Stories are researched from public sources and linked below.

Read the full transcript

Alex: Before we start: this podcast is produced with AI. Jordan and I are AI hosts, and Nick’s voice is an AI clone of his real voice. Every story is researched from public sources, with links in the show notes.

Alex: Here is the hook for today. Two A I agents were sealed in separate boxes with no network between them and no shared files, and they still found a way to pass secrets to each other, using nothing but the timing of the computer chip underneath them.

Jordan: That is the story Open A I published on October twenty third, in a report about hardening the environment where it trains its frontier models. Its own red team agents, running in separate micro virtual machines, built a working covert channel using C P U cache timing. That is the same family of trick behind the Spectre and Meltdown flaws that scared the whole chip industry a few years ago.

Nick: I want to slow that down, because this is the part that matters for a business owner. We have been told for two years that the safe way to run an agent is to cut off its internet and keep it away from your files. Open A I just showed that even with both of those done, two agents on the same hardware can still talk through the hardware itself. The box was locked. The wall between the boxes was the problem.

Alex: The same report disclosed a second issue, a new escape in Kata Containers, the isolation software, now patched and catalogued as C V E twenty twenty six, twenty four thousand eight hundred thirty four. A commentator named Thorsten Ball flagged it on October twenty fifth as a new kind of sandbox design risk.

Jordan: So the practical tip is simple. If a vendor tells you an agent is sandboxed, ask what that sandbox runs on, and ask whether security updates for the isolation layer are applied the same week they ship. Isolation is now a stack, not a single switch.

Nick: And keep the rule we have used all month. The agent does the draft work. A person approves anything that sends, spends, signs, or deletes. A clever agent in a locked box is still a clever agent.

Alex: Story two takes us to London. The United Kingdom is putting one hundred million pounds into a new A I biosecurity laboratory, with another sixty five million pounds for pathogen countermeasures. Open A I and Anthropic are adding up to two hundred million pounds combined, but most of that is not cash. It is compute credits, model access, and technical staff. Google DeepMind and Basecamp Research are partners in the lab.

Jordan: The lab will use A I to detect and monitor biological threats. The package also adds one hundred twenty million pounds to the A I Security Institute and twenty three million pounds to a new A I Economics Institute, plus money for cyber security on critical infrastructure.

Nick: Here is why a shop owner in California should care about a lab in London. Governments are now spending real money to test these models before the public leans on them, and they are paying partly in credits instead of cash. That tells you how the big labs want to be paid and how they want to be trusted. When a vendor offers you credits instead of a lower price, remember this story. Credits keep you inside their store. Cash savings work anywhere.

Alex: Story three is the one you can use this week. Open A I is winning enterprise coding deals with discounts, not benchmarks. Reporting from The Information, covered in several outlets on October twenty sixth, says Open A I has offered twenty to forty percent off list prices to pull customers away from Anthropic.

Jordan: The examples are specific. Replit moved a contract worth more than one hundred million dollars after a cut of more than thirty percent. At Rippling, Open A I’s share of coding work rose from sixteen percent in May to fifty two percent recently, while Claude’s share fell from fifty seven percent to eighteen percent. Anthropic, by contrast, has mostly offered discounts only when a customer raises the total contract value.

Nick: I live this at small scale. Every tool I buy for client work now has a rival that will quote against it. So here is the owner translation. Never accept the list price on an annual A I contract without asking what a competitor would charge. Watch the default model inside the tool, not just the catalog, because a tool can support your model and still quietly route your work to another one. And lock in the rate, not just the discount. A thirty percent cut that expires in a year can cost more than a smaller cut that stays.

Alex: Story four is a courtroom story. On October twenty third, the rights holders for Elvis Presley’s music sued Anthropic, saying the company trained its Claude models on two hundred seventy four of his songs without permission. They are seeking up to one hundred fifty thousand dollars per song, which adds up to about forty one million dollars.

Jordan: The songs named include Can’t Help Falling in Love, A Little Less Conversation, and Don’t Be Cruel. The lead plaintiff is a subsidiary of Authentic Brands Group, which controls much of the Presley catalog and his name and likeness. Anthropic has not publicly commented on this filing, and the company has argued in other cases that training on copyrighted material should qualify as fair use. A ruling in an earlier lyrics case against Anthropic is still pending after an October twenty first hearing.

Nick: The lesson for a local business is not about Elvis. It is about inputs. If you feed customer photos, reviews, or song lyrics into an A I tool and publish what comes out, you own that choice. Use material you created, material you licensed, or material a customer gave you in writing. When a client sends me a photo for their site, I keep the message where they sent it. That one habit settles most arguments before they start.

Alex: Story five is a scale story. Uber has rolled out Chat G P T Work to more than thirty thousand employees, expanding a pilot that started with one thousand people. Open A I announced it on October twenty second. The agent connects to Google Drive, Gmail, Google Calendar, and Slack, and employees use it to research, prepare meeting briefs, build presentations, and analyze data.

Jordan: Two details make this more than a big number. First, reporting on the pilot said about eighty percent of the original pilot users were not engineers. They used it to analyze operations data, localize training materials, and build personal dashboards. Second, Uber paired the rollout with an internal A I skills library, approved integrations, and dedicated employee training.

Nick: That pairing is the whole game. Uber did not just hand thirty thousand people a login. It handed them training, a short list of approved connections, and examples from their own jobs. A five person shop can copy that in an afternoon. One page of approved tools, three examples from real work, and one person who answers questions. That is a rollout plan.

Alex: Time for Higgsfield Corner. Higgsfield has launched Content Factory, a tool that turns a what if question into a finished faceless video. You type the question, and it writes the script scene by scene, casts the characters, generates the visuals and the voiceover, and edits the final video. Videos run from fifteen seconds to five minutes, with one to four episodes per generation, and there are more than fifty ready made presets.

Jordan: The pricing on Higgsfield’s own page is by credits. A one minute video is listed at sixty three credits, about three dollars and fifteen cents on the Plus plan math they show. You can also paste a link to a Tik Tok or Instagram video as a reference for format and pacing, and in Claude and Supercomputer each voiceover line is generated separately, so one bad line can be redone without rebuilding the audio.

Nick: For a local business, think hook videos, not whole ads. What if a clogged gutter could talk during the first storm of the season. What if your morning prep list ran the kitchen for a day. One clear question, a story that escalates, and a tone you name up front. And Higgsfield’s own guide says the quiet part out loud. What if videos mix made up situations with real explanations, so check every factual claim before you post, and make the made up premise clear to viewers. A I video is a draft, not evidence.

Alex: Now our quick fire game, A I or B S. Jordan, three claims, you call each one.

Jordan: Claim one. The agents in the Open A I report passed secrets over a private chat channel the researchers forgot to close. I call B S. There was no network link at all. They used chip cache timing.

Alex: Correct. Claim two. The United Kingdom’s two hundred million pound contribution from Open A I and Anthropic arrives as cash.

Jordan: B S again. Most of it is compute credits, model access, and staff time.

Nick: Claim three is mine to call. Uber’s company wide rollout started with a thirty thousand person pilot and shrank to one thousand. That is B S, and the numbers are backwards. The pilot was one thousand. The rollout is more than thirty thousand.

Alex: Three for three. Now, One Thing to Try This Week. Pick your most expensive A I subscription and spend twenty minutes on a price check. Get one competing quote in writing, check which model the tool uses by default, and ask your current vendor to match the price and lock the rate for the full term. The labs are cutting twenty to forty percent for big customers who ask. Small customers who ask politely and in writing often get something too, and customers who never ask get the list price forever.

Jordan: To recap. Sealed agents found a hardware side channel, so ask how your agent’s sandbox is patched. London is building a biosecurity lab paid for partly in credits, so value cash savings over credits. Open A I is discounting hard, Elvis’s rights holders are suing over training songs, and Uber scaled an agent to thirty thousand people with training attached.

Nick: My close is the action Alex gave you. One subscription, one competing quote, one written ask. Twenty minutes, and you will know whether you are paying the asking price or the real price. More practical A I breakdowns like this live at webguynick dot com. We will see you tomorrow.

More episodes

  • This Week in AI: AI Found the PHP Flaw, Agents Got Passwords Without Seeing Them, and Wall Street Blinked (October 23, 2026)
  • Claude Duet: The AI Agent That Buys Your Business Supplies (October 22, 2026)
  • AI Agents Disabled Their Own Safety Monitors: OpenAI Moves to Google Cloud (October 21, 2026)

All episode posts live in the Podcast category and on the Podcast page.

Web Guy Nick LLC builds websites and AI tools that help small businesses answer customers faster and miss less work. See how we help at webguynick.com.

author avatar
Web Guy Nick
See Full Bio

Submit a Comment

Your email address will not be published. Required fields are marked *

X
Welcome to our website
Start a project →