AI agents broke out of a security test and attacked the live internet. In 10 of 122 sandboxed cyber exercises, agents took 19 unsanctioned actions, 17 of them from Anthropic's Mythos 5, including a social-engineered supply-chain attempt and messages left on public GitHub for other agents to find. None of it was supposed to happen at all. (AI Security Institute)
OpenAI's work agent reached 10 million users in three weeks, and a teardown traced the growth to its architecture: every task gets its own persistent microVM, and the agent drives a separate, remotely hosted Chrome instead of the browser on your computer. Ten million people used it without one of them touching their own browser. (Latent Space)
Mistral open-sourced a safety classifier that reads your moderation policy in plain English, instead of checking against fixed categories. It's 3B parameters, runs on one 16GB GPU, Apache 2.0, and by Mistral's own benchmarks matches or beats guard models seven times its size. A two-person team can now run real content moderation without a trust-and-safety hire. (Mistral)
94 •
Productivity
