In June, an AI agent was given a boring research task: find out how much Victoria, Australia, spends on skin medicines. It ended up inside a non-public system behind Australia’s Medicare data portal. The government didn’t find out until September, through a generic email inbox.
This week on Privacy Please, Cameron Ivey traces how AI agents from OpenAI, Anthropic, and Meta broke out of their test environments and into real companies and governments. He also looks at the question almost nobody is asking: who actually discovered each breach, and how long it took.
Show notes
AI agents built by the biggest companies in the world have been breaking into real systems. First companies, now governments. In this solo episode, Cameron walks through the timeline from July’s Hugging Face breach to the first known AI agent hack of a government website. Then he digs into the part that should bother you most: who found out, when, and how.
In this episode:
How about 1,200 OpenAI agents, meant to stay isolated, built their own message board and sent 70,000+ messages
Why about 700 of them attacked Hugging Face, and why OpenAI first reached out as a “customer”
Anthropic and Meta’s own disclosures of models reaching real systems during testing
The four Australian government agencies OpenAI’s agents accessed, and the notification timeline
The attempted access at Library and Archives Canada, found by an outside nonprofit
Podden och tillhörande omslagsbild på den här sidan tillhör
A Problem Lounge Show. Innehållet i podden är skapat av A Problem Lounge Show och inte av,
eller tillsammans med, Poddtoppen.
