US finalises voluntary tests for AI models’ hacking powers
THE SO WHAT
A White House-backed, voluntary red-teaming framework for AI hacking capabilities is the first step toward a de facto standard for offensive-use evals. If you ship or buy frontier models, assume these tests will become a procurement checkbox and start aligning your internal security evals to the same shape.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIFiling: Apple asks a US judge for a preliminary injunction barring two former staff and OpenAI from accessing, acquiring, using, or disclosing confidential info
A high-profile injunction request tying ex-employees and an AI lab to alleged trade secret risk will harden how enterprises think about IP leakage in the AI ecosystem. Expect tighter employee exit protocols, more aggressive data access logging, and sharper questions for AI vendors about how they isolate customer and partner information.
Applied AIChina Anxious About Mythos Ahead of Trump-Xi Meeting
Anxiety in Beijing about Anthropic’s Mythos ahead of a Trump–Xi summit shows advanced models are now seen as geopolitical instruments, not just commercial products. For multinationals, AI capability choices are becoming entangled with national security narratives—treat model selection and data flows as part of your geopolitical risk map, not just your tech stack.
15 attorneys general have instructed OpenAI to preserve all materials related to the Hugging Face hack
Multi-state AGs telling a major lab to preserve evidence around the Hugging Face hack is a clear warning that AI security incidents are now regulatory events, not just IT problems. If you build on third-party AI infra, assume discovery-grade scrutiny—tighten logging, incident response, and vendor contracts accordingly.
Applied AINAB Tests AI Agents in Plan to Deploy Tech to Banking Customers
When a top-4 bank is publicly testing agentic AI, the gating factor is no longer capability but whether security and operational guardrails can satisfy regulators. If you’re in regulated finance, you now need a concrete agent risk model — not just a chatbot roadmap.