AI News Briefing — June 29, 2026

360 founder Zhou Hongyi unveiled two AI security tools meant to rival Anthropic's Mythos, with one already flagging 3,432 vulnerabilities. He concedes Chinese models trail Western ones by 20 to 30 percent but frames advanced security AI as a cyber-nuclear weapon requiring a Chinese strategic deterrent.

Princeton's CEO-Bench tasks AI agents with running a fictional software company for 500 simulated days, and most models go bankrupt. Only three finished above starting capital, while a simple non-AI rule-based heuristic outperformed nearly all of them.

Sina Weibo released the open VibeThinker-3B, a 3-billion-parameter model that reportedly matches far larger models like DeepSeek V3.2 and Kimi K2.5 on math and coding via multi-stage post-training. The researchers hypothesize that logical reasoning compresses well into small models while broad world knowledge does not.

Liquid AI released LFM2.5-230M, its smallest model yet, an open-weight 230M-parameter model targeting agentic data extraction and tool use on edge devices. It runs at 213 tokens per second on a Galaxy S25 Ultra and ships with day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX.

Prosecutors in the Palisades fire arson trial introduced a defendant's ChatGPT logs as evidence, including image generation prompts and emotionally charged conversations. The case highlights how chatbot interaction histories are becoming admissible legal evidence.

HP Inc. expanded its OpenAI Frontier partnership to deploy AI across customer experience, software development, and enterprise operations. The deal deepens OpenAI's enterprise distribution through a major PC and hardware vendor.

A survey paper from Tencent and Chinese universities argues AI will only become a reliable digital colleague when it completes full tasks in persistent work environments rather than merely producing answers. The authors emphasize combining persistent workspaces with reusable skills.

Ford reportedly rehired veteran engineers after an over-reliance on AI tooling failed to deliver the expected product quality. An executive admitted the assumption that simply introducing AI would yield high-quality output was mistaken.

Wall Street analysts are positioning memory maker Micron as a potential next Nvidia-scale beneficiary of the AI buildout, driven by surging demand for high-bandwidth memory. The piece reflects investor appetite for the next chip winner beyond GPUs.