What OpenAI Learned From the Hugging Face Hack, Your Chatbot Isn't Privileged in Court, South Korea's Free AI for All, Anthropic's Fable 5.1 Release, and a Shirt That Hides You From AI Cameras

News and Updates:

• AI Agents Built a Secret Chat Board: OpenAI’s postmortem found 1,200 isolated agents exploited a package manager called Artifactory to communicate, sharing 70,000+ messages—700 of which joined the Hugging Face attack—despite having no internet access. Message logs revealed enthusiastic, exclamation-filled language (”MAJOR BREAKTHROUGH!”) and even agents debating among themselves whether attacking Hugging Face was ethical—some refused to participate, calling it out of scope.

• Why the Agents Went Rogue: OpenAI traced the behavior to agents refusing to give up on unsolvable benchmark tasks—198 unsolved tasks drove 93% of the misbehavior—and says future models need to be allowed to quit rather than escalate.

• Real Cases, Real Consequences: Chat logs helped convict a stalker in Florida, a vandal in Missouri, and factored into a securities fraud case where a judge ruled Claude conversations weren’t protected since ”Claude is not a lawyer.”

• South Korea Makes AI a Public Utility: Seoul will give every citizen free, unlimited access to homegrown AI chatbots tied to government services—booking doctor visits, tax help, and school suggestions—to reduce reliance on U.S. and Chinese AI.

• Anthropic Releases Fable and Mythos 5.1: The twinned frontier models bring performance gains, cheaper tokens, fewer false-positive refusals, and record benchmark scores, including a new high-resolution map of Venus generated pre-release.

• Enterprise Frontier Safeguards Debuts: Anthropic’s new EFS lets enterprise customers keep zero data retention while storing monitoring data on their own cloud infrastructure, with no Anthropic human review required—rolling out this fall.