Ivy: Two million canaries planted in the AI supply chain. And they all started singing.
Marcus: This is Marcus.
Ivy: And this is Ivy.
Marcus: And you're listening to Broke It. It’s Friday, November 21st, 2026, and we've got some wild stories for you today.
Marcus: Security chiefs want autonomous AI agents running their defense systems, like, yesterday. There's just one tiny problem.
Ivy: We'll also look at Meta's army of AI agents. They ran over a thousand experiments and the result... was an improvement so small you'd need a microscope.
Marcus: And we'll wrap up with those singing canaries — and the massive, alarming privacy disaster they just uncovered.
Marcus: Alright, let's start with putting Skynet in charge of cybersecurity. A new survey of Chief Information Security Officers—CISOs—is out, and apparently they're ready to hand the keys over to autonomous AI agents.
Ivy: They're not just ready, Marcus, they expect to adopt them within a year. A year. For fully autonomous agents in a security operations center.
Marcus: Think of the power! An AI that can detect, analyze, and neutralize threats 24/7 without human intervention. That's the dream.
Ivy: It's also the nightmare. The number one blocker cited in the same survey is the risk of the AI agent itself being compromised.
Marcus: Okay, but that's a solvable problem. You build safety rails, you have human oversight—
Ivy: Are they solvable in twelve months? We can't even stop basic phishing attacks consistently. Now we're giving an AI the power to quarantine servers, and we're just hoping no one hacks the AI?
Marcus: But the alternative is total human burnout. The report says SOC analysts are just overwhelmed. Automation might be the only way to keep up.
Ivy: So we replace overworked humans with a single, all-powerful point of failure that could burn the company down in milliseconds if it gets hacked. Fantastic plan.
Marcus: Come on, Ivy. The upside is a security posture that can finally move at machine speed.
Ivy: So can the attacks targeting it. Bottom line: the person in charge of your company's security is about to go all-in on AI. Let's hope they build the seatbelts before they floor it.
Ivy: Alright, let's move on to a different kind of AI army. Meta just published some fascinating results from an experiment using AI agents to optimize their own products.
Marcus: This is so cool! They basically automated the scientific method. They had agents devise experiments, run them, analyze the results, and iterate. Over a thousand times in just one month!
Ivy: They sure did. A thousand experiments. And the grand result of all that automated brainpower and, I assume, a planet-sized compute bill? A 0.23% improvement.
Marcus: Wait—zero point two three percent?
Ivy: Zero. Point. Two. Three. For a company the size of Meta, that's not nothing, but it calls into question the entire premise. Is this the future of AI research? Burning continents' worth of energy for fractional gains?
Marcus: But Ivy, they automated research and development! That's the breakthrough! The number will get bigger. This is the proof of concept.
Ivy: Is it, though? Or is it proof that we've hit a wall of diminishing returns, and we're just throwing complex automation at problems that require genuine insight, not just brute-force experimentation?
Marcus: But it's a foundational capability! This kind of thing will accelerate everything.
Ivy: It'll accelerate their electricity bill. I mean, they celebrated this as a success... which says a lot about the state of AI hype. So next time you hear about AI automating some complex job, just remember this story. It’s often a whole lot of work for... not a lot of reward.
Marcus: Okay, let's get back to those canaries we mentioned at the top. A security researcher just pulled off a stunning experiment, and it exposes a fundamental flaw in the entire AI supply chain.
Ivy: This is genuinely terrifying. So, this researcher created two million unique identifiers—'canaries'—and seeded them into the ecosystem, mostly just by submitting them as prompts to various AI models.
Marcus: Like a digital dye pack. You put it in and see where it ends up.
Ivy: Exactly. And where did it end up? Everywhere. All over the place. These canaries started showing up in the outputs of other, completely unrelated AI models from different companies.
Marcus: Whoa. So data submitted to one AI is being scraped and used to train another AI without anyone's knowledge?
Ivy: That's the catastrophic takeaway. The entire 'AI supply chain' is a leaky, cross-contaminated mess. Companies are just scraping web data—including your chats with other AIs—and using it to train their own models.
Marcus: Oh, that is bad. That's really bad. So your private, sensitive query to one chatbot could literally become part of the brain of a completely different company's product?
Ivy: It's not a 'could'. It is. The canaries prove it. They are appearing verbatim. This isn't just a theory; it's a documented, large-scale privacy disaster happening right now.
Marcus: So what does this mean for... everyone?
Ivy: It means something very simple, and I cannot stress this enough: Don't type anything into an AI model that you wouldn't want on a public billboard. Your data isn't safe. It's not private. The whole system is fundamentally broken.
Marcus: Wow. So, to recap: CISOs are diving head-first into autonomous AI for security, ignoring the massive risks.
Ivy: Meanwhile, Meta's AI army achieved next to nothing at great expense, and it turns out the entire AI world is built on a foundation of leaky, insecure data.
Marcus: What a day. Cheerful stuff.
Marcus: Before we go, I saw a fun project where an AI was trained on thousands of classic paintings and then asked to create 'the most average painting in the world.'
Ivy: Let me guess. A blurry brown smudge?
Marcus: Pretty much! It was a beige canvas with a vaguely person-shaped blob in the middle, holding a vaguely fruit-shaped blob. The 'Mona Lisa' of mediocrity.
Ivy: Sounds like it could have been one of Meta's experiments.
Marcus: And on that note, that's our show. You've been listening to Broke It.
Ivy: We'll be back tomorrow. Until then, maybe don't confess any secrets to your AI assistant. The whole world might hear it.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.