Unitree Dex5-S dexterous hand offers 22 degrees of freedom

Nvidia argues AI security needs defined requirements, enforceable controls, named owners and evidence protections work, applied across models, harnesses and runtime environments. Example: an agent reading malicious instructions in an attached document should be blocked by network policy, with protected logs capturing the tool call and destination.

Openlayer's jevals swaps free-form LLM-as-judge scoring for typed Jev decisions in evals. Posted as a Show HN project on GitHub.

Hugging Face shipped tokenizers v1, a major version of its text tokenization library, with the release post focused on encode, decode, and scaling measurements.

V7 turns scattered company files into context agents can use for complex, source-linked work, built on GPT-5.6.

Guest article by Heather Fricke argues archived clinical data can encode old access patterns, reimbursement incentives, and unequal treatment that AI will faithfully learn without understanding why. She cites NIST's healthcare AI work on data quality and calls for adding provenance of meaning to governance.