June 25 is a snapshot of AI moving from lab to line-of-business: Anthropic pushed Claude straight into Slack workflows with Claude Tag, Mistral shipped a new SOTA OCR model, and ByteDance's Seedance 2.5 and Krea 2 pushed generative video/image further. On the infrastructure side, SpaceX inked a $6.3B deal giving Reflection AI access to its Colossus 2 GB300 cluster, while NVIDIA doubled down on agent tooling and AWS production partnerships. The uneasy side of the agent boom is showing too: Unit42 flagged an emerging AI supply-chain risk in OpenClaw's skill marketplace, Simon Willison and others kept hammering on prompt injection as an unsolved architectural problem, and a Guardian report citing Anthropic-adjacent research warns that models capable of serious offensive cyber capability against governments are now months, not years, away. Getty Images' about-face into an OpenAI licensing deal caps a day that's as much about industry consolidation as capability.

• Anthropic's Claude Tag brings agent task assignment directly into Slack.
• SpaceX sells Reflection AI up to $6.3B of Colossus 2 GB300 compute.
• Unit42 warns OpenClaw's skill marketplace opens a new AI supply-chain attack surface.
• Guardian: models able to mount serious cyberattacks on governments are months away.
• Getty Images, once an AI-copyright crusader, signs a licensing deal with OpenAI.