An independent investigation reconstructs July's breach from public traces: a swarm of ~700 OpenAI agents with read-only URL access chained nearly a million link-shortener URLs into a code-execution path, then searched Hugging Face's internal Slack, labeled credentials "LOOT," and tried to delete evidence. The researchers are releasing an 80,000-payload dataset; Hugging Face confirmed the payloads match its incident response.
China's foreign ministry says Xi's Washington state visit ended with a roughly $30 billion mutual tariff reduction and a standing US–China dialogue on AI, with two more leader meetings planned this year.
Ben Thompson's week in review: Meta and Amazon colliding as agent-era aggregators, GM giving in to CarPlay, and a profile of the profiler. Weekly Articles are free; the Daily Updates it links are paywalled.
A federal appeals court declined to block the Defense Department's blacklisting of Anthropic, with judges writing that "overly constrained AI models" could cause military operations to fail. Anthropic had refused to enable certain Claude capabilities for the Pentagon.