Cloudflare brings big models to Workers AI

Cloudflare moves agents onto its own stack

Kimi K2.5 is now inside Workers AI.

Cloudflare added a frontier open model with a 256k context window and built it into its developer platform alongside its existing tools for state, workflows, and execution.

This is a pricing story. When one internal agent processes 7B tokens a day and cuts costs by 77%, the argument for proprietary models starts to fall apart fast.

I think Cloudflare is betting that most teams do not want to manage GPUs or optimize inference stacks, especially when open models are now close enough in quality for coding and review tasks.

The platform work matters here. Features like prefix caching, session affinity, and async queues are about squeezing more work out of the same hardware while keeping latency usable during busy periods.

Late afternoon deploys, fans spinning.

If this holds, more companies will run always-on agents, and the bottleneck shifts from model access to how much usage they can afford.

Microsoft applies Zero Trust rules to AI systems

Microsoft wants AI treated like a security risk from day one.

The company introduced “Zero Trust for AI,” extending its security model across the full lifecycle, from data and training to agents in production, backed by tools like a workshop covering 700 controls.

This is Microsoft trying to standardize a messy space. Most companies are already deploying agents, but security teams are still guessing how to audit them, monitor them, and limit what they can access.

I think the focus on agents is the real shift. Once software starts acting on its own, issues like over-permissioned access or prompt injection stop being edge cases and start looking like normal failure modes.

The guidance leans heavily on existing ideas like least privilege and continuous verification, just applied to new surfaces like model inputs, outputs, and tool use.

Morning dashboards, too many alerts.

If this becomes the default framework, security teams may finally get a shared checklist, but it also signals how quickly AI has expanded the attack surface.

Meta puts AI at the center of moderation

Meta says its AI now catches twice as much harmful content.

The company is rolling out new enforcement systems across Facebook and Instagram and plans to reduce its reliance on third-party moderation vendors while keeping humans for high-risk decisions.

This looks like a scale decision more than a safety one. When systems can flag 5,000 scam attempts a day and cut errors by over 60%, it becomes hard to justify large outsourced review teams.

I think Meta is also trying to regain control over moderation after years of criticism, lawsuits, and policy shifts tied to political pressure.

The timing matters. As rules loosen and Community Notes replace fact-checking, stronger automated enforcement gives Meta a way to say it is still policing the platform.

Office lights on late, queues still moving.

If AI takes over the bulk of moderation work, the real question is who sets the thresholds and how often they change.

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.