Trends in AI development
Agent Skills now has a neutral home at the Linux Foundation. The format matters less than what teams put in it: written procedures that people and agents can both check.
Read articlePrompt injection can't be patched out of AI agents. The October 2026 evidence points to a plainer fix: limit what each agent can read, change and send, and keep humans on the risky actions.
October brought a wave of large open-weight models. Before picking one, look at the license, the hardware it needs and how it does on your own tasks.
Soft spend alerts are not enough when agents can call paid APIs overnight. Default hard budget caps are becoming basic infrastructure for AI builders.
