3 numbers from this week's AI news that should be on every business owner's...

3 numbers from this week's AI news that should be on every business owner's radar.

88% of companies running AI agents have had a security incident. Only 23% have any agent-specific security framework.

Nearly nine out of ten companies deploying AI agents have already had something go wrong, and fewer than one in four have built any safety net for those agents. That's not a typo. In March, a single compromised software package sat in a public repository for three hours, racked up 47,000 downloads, and infected agent frameworks at multiple major companies... running autonomously the entire time, no human attacker needed.

AI agents score 59% accuracy on real workplace documents.

Not toy benchmarks. Real workplace files... messy, inconsistent, the kind of documents your business actually runs on. Researchers tested eight leading models against 362 questions over nearly 10,000 authentic business documents. The best model got 59.4% right (and that was the best model). That's the agent you're trusting to summarize your contracts, pull numbers from your financials, and draft your compliance reports.

A German court just ruled that AI errors are the company's liability. Not the model's. Yours.

Google got hit with a ruling that their AI Overview answers are Google's own words... and Google is liable for mistakes in them. Bruce Schneier put it plainly: if you'd be liable for a human employee making the same mistake, you're liable when your AI makes it. This will be the template. Every company using AI with customers should be asking: can we survive the first lawsuit?

Here's the pattern. Companies are shipping AI agents at full speed with almost no safety infrastructure, the agents themselves are less reliable than the sales pitch suggests, and courts are starting to hold companies accountable for what their agents do.

The hype says "AI agents are here." The data says "AI agents are here, and most of you aren't ready."

If you're running agents (or about to), start with two questions: what can they access, and what happens when they're wrong?