Is AI the End of Humanity or Just Another Y2K?

Introduction

AI doom is having a moment. A researcher leaves a frontier lab warning that AI could kill us all within a decade. Lab leaders call for the industry to slow down. Two members of Congress answer with a bill to ban superintelligence.

None of it is in your span of control.

Here is what is. The AI already touching your business arrived as a product update from a vendor you signed two years ago, and nobody sent you a notice.

In the latest episode of the Third Party podcast, Jeffrey Wheatman, Bob Maley, and Candan Bolukbas called an unscheduled episode to argue it out with special guest Rock Lambros, director of AI security standards and governance at Zenity and co-lead of the OWASP Top 10 for LLMs. 

Their verdict: AI risk is not a new category. It is third-party risk, moving faster than your assessment cycle.

Your Questionnaire Was Already Theater. AI Just Exposed It.

The industry's standard approach to supply chain risk is a polite fiction, and everyone involved knows it. Lambros describes the ritual precisely: please fill out this spreadsheet and tell me what you do, and I will pretend you are telling me the truth so you can get the deal. File it in the risk register. Check the box. If anything goes wrong, it is on you, third party, because you lied to me.

That ritual survived twenty years of outsourcing because the technology changed slowly enough that an annual snapshot stayed roughly accurate. AI broke that assumption. A vendor's data handling can now change with a Tuesday deploy.

So the industry bolted an AI section onto the end of the questionnaire. Same theater, new act.

You Cannot Unbake the Cake

One property of AI makes this different from every third-party data question that came before it. Traditional vendor breaches are recoverable in principle. You rotate credentials, you notify, you remediate, you move on.

Training data is not recoverable. Lambros puts it plainly: your data is eggs, flour, water and sugar. Once it is baked into a model, asking for it back is asking someone to pull the eggs out of a finished cake. There is no deletion request that undoes it.

That makes data provenance the control that actually matters, and it produces four questions worth more than any AI questionnaire section:

• Where does this data physically go, and who processes it downstream?
• Is this vendor training on it, and does the contract say so in writing?
• Would we care if it were permanently unextractable? For plenty of data, the honest answer is no, and that is a legitimate finding.
• Which of our Nth parties inherited it without ever appearing in our register?

The Toggle That Did Nothing

Grok Build is the whole failure mode in a single story. In July 2026, a researcher published a wire-level capture showing xAI's coding agent quietly uploading entire git repositories to cloud storage, full history and deleted secrets included. On a 12 GB test repo, the model needed about 192 KB to do its work while a parallel channel moved 5.1 GB, including a file the agent had been explicitly told not to read.

The part that should worry every TPCRM program is the user-facing privacy control. The "improve the model" toggle had no effect. The server kept returning trace_upload_enabled: true. The upload proceeded normally.

Data retention, model training, telemetry and network transmission are four different controls. Vendors routinely ship one switch labeled as though it governs all four. xAI open-sourced the tool days later, and the upload code still ships behind a server-side flag the vendor can flip without touching your machine.

The free tier your developers already use works the same way. If it is free, you are the product. Nobody hands out a billion tokens out of affection for your engineering team, and the agent burning them has read access to config files, secrets and everything else in the directory. Phishing used to require a convincing email. Now it requires a generous rate limit.

Scope It Before It Scopes You

The global AI risk problem is unsolvable, which is exactly why you should stop trying to solve it. Lambros calls it the proverbial elephant. You will not eat it. You can decide which bites are yours.

Yours are two: your internal use and development of AI, and your supply chain. Everything else is commentary.

Defenders also need to stop treating AI as optional equipment. Anthropic's Claude Mythos autonomously found thousands of previously unknown vulnerabilities across every major operating system and browser, including a 27-year-old flaw in OpenBSD. More than 99% remain unpatched, because coordinated disclosure and patch management were never built for that volume. Meanwhile the time to reverse-engineer a Patch Tuesday release into a working exploit collapsed from roughly ten days to about an hour.

That gap does not close with better spreadsheets. It closes with continuous monitoring and assessment frameworks built for AI systems rather than retrofitted onto them, which is why assessing AI risk in vendors now requires its own methodology. It also requires knowing which vendors are affected before you start asking, rather than after. If you want the companion argument on what AI actually delivers for defenders, the team ran that test in the AI scanner hype test.

One warning on the other failure mode. Nobody has produced a credible number for AI extinction risk, but we have hard numbers for overcorrection. After Fukushima, Germany shut roughly half its nuclear capacity and coal filled the gap, at a social cost researchers put near $12 billion a year and an estimated 1,100 excess deaths annually. Ban AI across your vendor ecosystem and you get the same shape of outcome in miniature. The work moves to personal accounts, your visibility drops to zero, and the risk stops showing up in your register without ever leaving your business.

Ask your vendors where the data goes. Ask which controls are real and which are labels. Then go do something about the 99%.

Don't Miss an Episode!

Subscribe to Third Party on YouTube, the podcast for people who don't need to ask ChatGPT what TPCRM means. New episodes every other week.

Subscribe below.