Skip to content
Inflection Point

The attackers didn't want the chatbot.

Your crown jewels list was probably written before AI. It names the customer database, the finance system, maybe the source code.

They wanted what it could read.

Here's what happened.

Drift is an AI chat agent. Companies put it on their websites and connect it to Salesforce, so it can see customer records. That connection runs on login tokens.

In August 2025, attackers stole those tokens. From August 8 to at least August 18, they used them to walk into customers' Salesforce accounts and pull out records: accounts, user records, sales deals, and support cases.

A Google threat analyst said more than 700 organizations may have been affected. Zscaler and Palo Alto Networks confirmed they were among them.

Then, Google says, they searched the stolen data for AWS keys, passwords, and Snowflake tokens.

Think about that for a second.

Why would keys to other systems be sitting in a CRM? In my experience, support tickets are a common place. When a customer is stuck, your support team pastes whatever it takes into the ticket. Do that for years and you've built a record of how your company works. Some of it is keys.

Salesforce itself wasn't broken into. The attackers came in through a connection somebody had approved.

Here's why this matters to you.

Now Salesforce is telling its customers to build more of that record, on purpose.

In September, Salesforce president Rohan Kumar laid out where he thinks the value in AI is going. The agent software, and where it runs, will become a commodity, he told diginomica. What Salesforce plans to sell is, in his words, "trusted enterprise context."

By context he means your data plus an understanding of how work actually gets done at your company. Much of that lives in meeting notes, email, and tickets. The pitch is to gather it up and organize it so AI agents can read it.

He may be right. Keep in mind that he sells that layer.

And you don't have to plan it for it to happen. Your people connect AI tools to email, shared drives, and the CRM. Each one starts reading that record the day it's connected.

Here's the mechanism.

Context used to be scattered. A decision lived in one email. A workaround lived in one ticket. A pricing exception lived in one person's head.

That was a pain. It also slowed thieves down. They had to find it piece by piece.

Organizing it for agents removes that protection. Whatever makes it useful to an agent makes it useful to a thief.

And the way in usually isn't a hack. It's a connection someone approved. In April, Vercel traced a breach back to a third-party AI tool one of its employees used. The attackers compromised the tool, then used that access to take over the employee's work Google account.

Nobody checking what the connected apps can read is a Basic Persistent Failure (BPF). It's an old weakness with new stakes.

Your company isn't Palo Alto Networks. It doesn't have to be. Do you run a CRM, a help desk, or Microsoft 365 with apps connected to it? Then you have the same setup on a smaller scale.

So here's the reframe.

Your crown jewels list was probably written before AI. It names the customer database, the finance system, maybe the source code. It doesn't name your context layer, because a year ago you didn't have one.

Put it on the list.

That doesn't mean lock it in a vault. The context is where the value is. Lock it up and your agents are useless, and your people will route around you.

It means you can answer three questions, the same ones you'd answer for any crown jewel.

Who owns it? Not IT by default. The business leader whose decisions are in it.

What can read it? Every agent and connected app, by name.

What's in it that shouldn't be? Passwords in support tickets were Drift's lesson.

If you can answer those, you can say yes to more AI with a straight face.

Who at your company owns the context your AI tools read?

-Kip