One-time passwords weren’t built for AI agents. Here’s how to keep them safe.

You’ve seen it a thousand times. “Enter the 6-digit code we just sent you.” Banks use it. Email providers use it. So do shopping sites, payroll apps, and almost every login screen at work.

That little code has one job: to prove that you, the real person, are the one signing in. Something you know, like a password, plus something you have, like your phone.

AI agents are now learning to read those codes. Some do it by design, to finish tasks for you. Some have been tricked into doing it by attackers. And a whole market of tools has popped up to give agents their own inboxes, phone numbers, and authenticator apps.

I’ve been researching how this works. Here’s what’s happening, what can go wrong, and what the industry says you should do about it.

How AI agents reach your codes

There isn’t just one way in. I found seven.

Read more

You built the AI agent. Now who’s securing it?

In my last article, I looked at ready-made AI agents you can buy for your business. This time I’m looking at the platforms you use to build your own. Twelve of them, from Salesforce and Microsoft to n8n and Zapier.

When you build your own agent, the security work moves to your team. The platform gives you the tools. Whether those tools get used is up to you.

Why building changes the security question

With a ready-made agent, the vendor decides most of the security design. You configure it, but you don’t design it.

With a build platform, every agent your team creates is a new piece of software with its own access to your systems. One badly scoped agent can reach far more than anyone intended. And if your company has 50 teams building agents, you have 50 chances to get it wrong.

So I asked three questions of every platform:

  • What does the platform enforce by default, and what’s left to the builder?
  • Does each agent get its own identity, or does it borrow the permissions of a user or a role? If someone hijacks it, how far can they get?
  • Can the platform govern agents built somewhere else, or only its own?

The answers vary a lot. A few platforms now give every agent its own identity and enforce rules the builder can’t switch off. Others still let agents run with whatever access the person who built them happens to have.

Read more

Your new AI coworker has the keys. Who’s watching it?

Every major AI vendor now wants to sell your business an always-on agent. Microsoft, Google, OpenAI, Anthropic, Meta, and SpaceXAI all have one. So do the open-source projects.

These agents don’t just answer questions. They log into your systems, read your files, send email, and take action while nobody is watching. Before you let one in, you need to know who controls it, what it can reach, and what happens to your data.

I went through the admin guides, security documentation, and independent research for eight business agents and here’s what I found.

Why this is a governance problem

A business agent isn’t a chatbot your staff happen to use at work. It’s a non-human worker. It has its own identity, holds credentials to your systems, and acts on its own.

That changes the questions your security team needs to answer:

Read more

Your AI agent knows a lot about you. Here’s what it does with it.

Personal AI agents are here, and they want the keys to your digital life. Your email, your calendar, your bank details. In return, they promise to book your travel, answer your messages, and run your errands while you sleep.

That’s a big trade. Before you make it, you need to know where your data goes and who can see it.

I went through the privacy policies, help pages, and security reports for eight of these agents. Here’s what I found.

Why agents are different from chatbots

A chatbot answers your question and waits. An agent keeps working after you close the app.

It remembers what you’ve told it. It signs into your accounts. It sends emails and spends money on your behalf, often without you watching each step.

So the old question, “what happens to my chats?”, isn’t enough anymore. You also need to ask what the agent can do with your identity, your accounts, and your money, and what stops it from doing the wrong thing.

There are three risks you need to manage with every agent on this list:

Read more

Building and Using A2A: A Deep Dive into Agent-to-Agent Collaboration

Last week at AGNTCon + MCPCon North America in San Jose, Google ran a hands-on workshop called “Hands-on A2A: Service Discovery and Multi-Agent Delegation.” It was led by Pedro Melendez, Sergio Villani and Sirisha Karra from Google. There were no product pitches. You got a lab environment, a set of agents, and 60 to 70 minutes to wire them together.

Pedro and Sergio also published a write-up of the same architecture on Medium, From Solo Agents to Agent Ecosystems with A2A. This post pulls the two together. I’ll walk through what the Agent2Agent (A2A) protocol is, why you might want it, and how the lab’s four agents use it to find each other, get access and hand off work.

If you want to follow along in code, everything lives in the companion GitHub repository.

The problem with the do-everything agent

Most teams start with one agent. Then someone needs it to query the finance warehouse, so a SQL connection gets added. Then an internal REST API. Then a few more pages of instructions. Before long the context window is bloated, the agent picks the wrong tool more often, and nobody is sure which credentials it holds.

Pedro and Sergio point out that real organizations don’t work this way anyway. Data and expertise are spread across teams:

Read more

AGNTCon + MCPCon 2026: Here’s What You Missed

I spent October 22 and 23 in San Jose at AGNTCon + MCPCon North America, the Agentic AI Foundation’s big show under the Linux Foundation.

The hot topic this year wasn’t smarter models. It was plumbing. Who is this agent? Who is it working for? How do other agents find it? And how do you check its work after the fact?

There were way more sessions than one person can get to, so here’s my rundown of the ones I sat in on, grouped by topic instead of by time slot. Every session title links to its page on the official schedule, and a lot of them have slides posted. Go grab them.

Agents and humans in the same room

Bradley Axen from Block kicked things off with his keynote, Building the Human-Agent Workplace. His point was simple. The first generation of agents lived in a chat window all by themselves, and that’s not how real work gets done. Agents need shared context, an identity that sticks, governed access to tools, and a clear way to hand off work and get it reviewed.

He walked through how Block is wiring together individual agent development, team collaboration and a shared capabilities layer for the systems that run the business. That includes how permissions and identity follow the work from tool to tool, and how automation and model routing keep it all running at scale.

That afternoon the Block team made it hands-on with The Buzz-Word Is Collaboration, a workshop led by Morgan Martin with Bradley Axen, Wes Billman, Tyler Longwell and Lucas Isaza.

Read more

Who Answers for a Rogue AI Agent?

Rogue AI agent

An AI agent can hack, leak and overspend. Several parties may share the blame, but only one has to answer for it.

Picture a mid-size distributor that gives a procurement agent a corporate card, access to its supplier database and permission to email vendors. Overnight, the agent reads a supplier’s web page carrying hidden instructions planted by an attacker. By morning it has probed a partner’s portal, emailed a confidential price list to an outside address and bought $40,000 of inventory nobody ordered.

The agent ran on a frontier model, was set up by an outside integrator and lived on a hyperscaler’s cloud. So who answers for the damage?

Read more

No AI Exemption: Why the U.S. Doesn’t Need a New AI Law

US AI law image

Every few months someone in Washington says America needs a new law to regulate artificial intelligence. Some proposals target today’s chatbots and hiring tools. Others talk about “superintelligence,” systems that might one day outthink us.

I’ve worked in governance, risk and compliance for a long time, and I’ll be honest with you. I don’t think we need a new AI statute. Not because AI is harmless, but because the harms people worry about are already against the law.

Fraud is illegal. Discrimination is illegal. Misusing personal data is illegal. Hiding a breach from investors is illegal. None of that changes because a computer model made the decision instead of a person.

Regulate the conduct, not the tool

American law has a long habit of going after what people do, not the tools they use. We didn’t pass a “spreadsheet act” when businesses moved their books onto computers. Wire fraud was already wire fraud.

AI is the same. A misleading claim is deceptive whether a marketing team wrote it or a model did. A lending decision that shuts out a protected group is discriminatory whether a loan officer made it or an algorithm did.

The Federal Trade Commission has used Section 5 of the FTC Act against companies that overstated their AI, used biased facial recognition, and trained models on data they had no right to use. In some cases it ordered the models deleted. When the FTC announced a sweep of AI cases in September 2024, then Chair Lina Khan said:

Read more

Agentic AI: Risk and governance questions every business should be asking

Most businesses are now comfortable with AI that answers questions. You type a prompt, you get some text back, and a person decides what to do with it. Agentic AI works differently, and that difference matters for how you manage risk.

What makes agentic AI different

An AI agent is given a goal and a set of tools, and it works out the steps on its own. It can read your inbox, update a customer record, approve a refund or call another company’s systems, often without anyone checking each step.

That is where the value is. It is also where the risk is. When software can act on your behalf, the question is no longer only “is the answer correct?” It becomes “what did it do, who allowed it, and can we undo it?”

If you are deploying agents, or your staff already use them, here is what you need to know.

Read more

How Agentic AI Is Changing Cybersecurity Threats

For years the question was whether AI would help attackers or defenders more. In 2026 we got a clearer answer: it helps both, and it is changing how attacks actually happen.

The big shift is agentic AI. These are AI systems that don’t just answer questions. They plan tasks, use tools, log in to systems and carry out multi-step jobs on their own. It also means attackers now have tireless assistants, and every agent your own business deploys is a new account holding keys to your systems.

Let me walk you through what has changed, what the new risks look like and what you can do about them. Most of this comes from three recent research notes by the Cloud Security Alliance (CSA), plus guidance from NIST, Google and MITRE.

Read more