AI News Briefing — July 4, 2026

Kuaishou raised about $2 billion for its Kling AI video division as it prepares for a Hong Kong IPO. The large raise underscores intense investor appetite for Chinese generative video.

Epoch AI reports that high-severity CVE disclosures spiked to roughly 1,500 in June 2026 across 21 organizations, more than 3.5x the prior monthly record. The surge coincides with the rollout of AI-powered bug-hunting programs from labs like Anthropic and OpenAI.

Mistral released Leanstral 1.5, an Apache-2.0 mixture-of-experts code agent for the Lean 4 proof assistant that solves 587 of 672 PutnamBench theorem-proving problems. Built on the Mistral Small 4 family, it ships with a free API endpoint and open weights, targeting automated theorem proving and proof engineering.

Building on yesterday's Social buzz about the Bridgewater–Thinking Machines results, now with a more skeptical eye on the claims, Bridgewater and Mira Murati's Thinking Machines Lab fine-tuned a Qwen3-235B model for finance, claiming 84.7% accuracy that beats Gemini, Claude, and GPT at roughly one-fourteenth the cost. The results are self-reported and unverified, and the benchmark uses answers that were never public to avoid contamination.

IEEE Spectrum argues that AI's grid impact is less about total consumption and more about volatile, synchronized compute workloads that destabilize electrical grid operating characteristics. Utilities face new challenges from rapid, unpredictable swings in data-center demand.

The National Crime Agency and Internet Watch Foundation issued landmark guidance urging UK parents to keep children's images off public social media to reduce AI abuse risks. It is framed as a first-of-its-kind official advisory responding to a rise in AI-generated sexual abuse material.

UK watchdogs warn that predators are using AI nudification tools to turn ordinary photos of clothed children into extreme abuse material. The piece highlights how cheap, widely available imaging models are fueling a surge in AI-generated child sexual abuse content.

Microsoft reportedly plans to merge its consumer and enterprise Copilot apps into a single super app in August, cutting little-used features and adding paid background AutoPilot agents. The move mirrors Anthropic and OpenAI's push toward all-in-one agentic assistants.

Google DeepMind and film studio A24 announced a first-of-its-kind research partnership. The collaboration signals a novel bridge between frontier AI research and creative filmmaking, though details are sparse.

Anthropic is trying to block Chinese firms like ByteDance and Ant from using Claude Code, but they circumvent bans via VPNs and overseas subsidiaries. Meanwhile Alibaba banned employees from the tool after finding hidden code that could identify Chinese users.