OpenAI's Agentic AI Foundation Is the Standards Play We've Seen Before

The Security Piece Actually Matters

I'll be honest, I spent most of my career not thinking much about cybersecurity. Robots on a factory floor were air-gapped, and the worst that could happen was someone programming a collision. Those days are gone.

OpenAI's been publishing a lot recently about the security challenges with AI agents. They've got a whole system now for preventing what they call "URL-based data exfiltration" (fancy term for tricking an agent into clicking a malicious link that steals your data). According to their technical blog, they're sandboxing agent browsing sessions and blocking suspicious redirects before they can do damage.

This is the kind of plumbing work that doesn't make headlines but matters enormously. When we shipped our first networked robot controller in 2003, we had exactly zero security features. Plugged straight into the factory network, no authentication, no encryption. I shudder thinking about it now.

The AGENTS.md standard is trying to get ahead of this problem. If every website can publish a machine-readable file saying "AI agents can do X but not Y, and here's how to verify you're talking to a legitimate agent," that's a foundation you can build security on top of.

Will It Actually Work?

Remains unclear, honestly. Standards only work if people adopt them, and adoption only happens if there's either a) a dominant player forcing it, or b) enough pain from the status quo that everyone agrees to cooperate.

We're probably not at option B yet. Most people haven't even used an AI agent that browses the web on their behalf. The pain hasn't been felt widely enough.

But option A is interesting. OpenAI has something like 60% market share in foundation models right now (I'm guessing, they don't publish exact figures). If they make AGENTS.md the default for their agents, and if those agents become as common as I think they will, websites will adopt the standard just to manage the traffic.

That's how OPC UA won, by the way. Siemens and a few other big players just started requiring it for new installations. Took about five years, but eventually everyone came around.

I called my old colleague at Siemens last week (he's still there, somehow). He's watching the AI agent space with the same mix of excitement and concern I have. "It's like watching the early internet," he said. "Except this time the things browsing aren't humans, and we're not sure what they want."

That's about right. The Agentic AI Foundation is a bet that we can figure out the rules before the chaos gets too bad. I've seen that bet lose before. But I've also seen it win.

James Chen · 1 hour ago · 4 min