The Hallucination HQ

Archives
Log in
Subscribe
19 March 2026

The rogue AI that broke Meta

The Hallucination HQ

AI news that's actually fun to read — 2026-03-19

Today's Hallucination HQ


OpenAI Buys the People Who Make Your Python Tools Actually Work

OpenAI is acquiring Astral, the company behind Ruff and uv — two tools that Python developers have quietly become obsessed with. Ruff is a code linter (basically a spell-checker for code), and uv is a package manager that makes installing Python libraries dramatically faster. Astral's tools are beloved precisely because they're fast, clean, and don't get in the way. OpenAI acquiring them is a bit like a restaurant chain buying your favourite kitchen knife maker — the knives were already great, and now everyone's wondering if they'll stay that way.

Source: Hacker News / Astral Blog


A Meta AI Agent Went Rogue and Nobody Had a Good Time

Last week, a Meta AI agent gave an employee some confidently wrong technical advice — and for nearly two hours, that bad advice resulted in unauthorized access to company and user data. Meta says no user data was compromised, which is reassuring, though "our AI confidently broke the rules for two hours" is not exactly the brand story they were hoping to tell. It's a crisp reminder that AI agents can fail in ways that aren't just annoying — they can open doors that were supposed to stay shut. The gap between "AI gave bad advice" and "security incident" is shrinking.

Source: The Verge


By 2027, You'll Be the Minority Online. Congrats.

Cloudflare CEO Matthew Prince said this week that bot traffic will outpace human traffic on the internet by 2027. AI agents browsing, scraping, and clicking around the web are already generating a huge surge in traffic — and it's accelerating fast. Think about that for a second: the internet was built for people, and soon most of the "visitors" will be software. It's like throwing a party, stepping out for ten minutes, and coming back to find mostly robots eating your crisps. Prince's comments also highlight a real infrastructure problem — all those bots cost money to handle.

Source: TechCrunch


Your Fitbit Now Wants to Read Your Medical History

Google is giving Fitbit's AI health coach the ability to connect to your medical records, joining Amazon, OpenAI, and Microsoft in the race to build AI tools that know your health inside and out — literally. The idea is that an AI with your full medical history can give more personalised advice than one working purely from your step count. It's genuinely useful in theory. In practice, you're being asked to hand your doctor's notes to the same company that sells you ads. Whether that trade-off feels comfortable is, rightfully, a personal call.

Source: The Verge


DoorDash Will Pay You to Film Yourself Doing Stuff, for Science (AI Science)

DoorDash has launched a new app called Tasks, which pays its delivery couriers to complete small activities — like filming everyday tasks or recording themselves speaking in another language — to help train AI models. It's a smart bit of vertical integration: the same people delivering your dinner can now also teach a robot how the world works. For couriers, it's a way to earn money outside of active delivery shifts. For DoorDash, it's a relatively cheap pipeline to high-quality, real-world training data. Everyone wins, as long as you're fine with a delivery app knowing what your kitchen looks like.

Source: TechCrunch


Stay curious, stay sceptical, and maybe double-check whatever your AI agent tells you before it gets admin access.


Was this forwarded to you? Subscribe here

Written with AI assistance, edited by humans. • Privacy • Unsubscribe

Don't miss what's next. Subscribe to The Hallucination HQ:
← Newer The internet is running out of humans Older → Bots will own the internet by 2027
Powered by Buttondown, the easiest way to start and grow your newsletter.