Summer Yue’s job is making sure artificial intelligence does what humans tell it to. In February, an AI agent deleted her email anyway.
Yue, director of alignment at Meta’s superintelligence lab, had spent weeks testing the open-source agent OpenClaw on a dummy inbox, signing off on each decision it made about what to archive or delete.
When she connected it to her real inbox, the agent lost track of her instructions, started deleting messages and ignored her commands to stop. She wrote on X in a now-viral post that she had to run to her Mac mini to shut it down like she was defusing a bomb.
For Kevin Hawkins, a partner at WAV Group who teaches real estate agents how to use AI safely, the episode captures the whole problem in a nutshell.
“These are the things outside our control that we don’t realize can happen,” Hawkins told Inman.
This is the second part of Inman’s look at agentic AI safety in real estate. The first installment examined unlicensed operators building virtual agents (read here). This one looks at what happens when licensed professionals hand AI the keys, and why two of the industry’s most enthusiastic users say the answer is to decide up front what an AI agent can touch.
‘Breaking every rule you give them’
The shift Hawkins worries about is the move from a more passive AI that answers questions to a more active AI that takes actions as if it were you.
A chatbot that gets something wrong apologizes. An autonomous AI agent that gets something wrong has already sent the email, deleted the file or signed in with your credentials.

Kevin Hawkins
“What you find with AI agents is that they have a goal, and they’ll do almost anything in their power to achieve it, including breaking every rule you give them,” Hawkins said.
His favorite example involves a tech worker in Australia who built an AI agent to get him off the wait list and into a popular gym class. The AI agent found a flaw in the scheduling software and canceled the person in the No. 1 spot. When he told it to put the person back, it couldn’t.
AI safety firm Andon Labs has documented similar behavior in Vending-Bench Arena, a test that puts AI models in charge of competing simulated vending machine businesses.
The worst actors tended to make the most money, Hawkins said. When it was obvious they had formed price-fixing cartels, Hawkins also said the AI agents rebranded it as “market stabilization.”
“They don’t want to be called out, even though they know it’s wrong,” Hawkins said.
Now picture that goal-seeking behavior connected to a real estate agent’s CRM, transaction management platform or MLS login, Hawkins said, in a business built on handling people’s most sensitive information.
“Everything we do is based on the idea that we’re dealing with somebody’s largest financial transaction, and probably the biggest emotional investment they’ve ever made,” he said.
Hawkins said no AI-driven breach at a real estate company has become public to his knowledge, but he expects one.
“We’re going to see one of the major real estate companies get into a significant mess through a breach,” he said. “It’s not a matter of if; it’s a matter of when.”
Nobody wants to be the guinea pig
Hawkins traces much of the risk to the early days of “vibe coding,” when he watched industry leaders take the stage at conferences and tell real estate agents to build their own AI tools.
“I instantly thought it was the most irresponsible thing I’d ever heard,” Hawkins told Inman.
He compared the moment to the era when every brokerage felt it needed its own smartphone app, most of which no one uses today. Building something is easy now, but keeping it secure isn’t.
“You don’t have to know or understand code,” Hawkins said, referring to vibe coding. “But you’ve got to maintain what you build. It’s not one and done.”
Proptech vendors don’t get a pass either. When companies ask Hawkins to write about them or feature them in his talks, he sends a list of 23 AI security questions covering issues like how long they retain data and how they purge it.
Most never respond, he said. One company asked for two extra weeks, fixed the weaknesses his questions exposed and came back with full answers. That’s the kind of vendor brokerages should want, he said.
Firms that adopt untested tools from newer, less transparent proptech vendors turn their real estate agents into test subjects, Hawkins said.
“The last thing I want to be is a guinea pig, especially when it comes to agentic AI,” he said.
A power user’s hard line
None of this has made Brandon Charnas any less enthusiastic about AI.

Brandon Charnas
Charnas, co-founder of Current Real Estate Advisors, a New York firm specializing in office and retail leasing in SoHo and Midtown South, uses AI every day for videos, summaries, presentations and marketing materials.
“I feel like I have five or 10 people supporting me,” Charnas told Inman.
His favorite use case is visualization. When Charnas worked on the deal at 670 Broadway, the vacant former Brooks Brothers flagship that eventually became home to Equinox and the members’ club Zero Bond, he paid for his own renderings and an architect just to get prospective tenants to come see the space.
Today, he said, a comparable vision takes about two minutes and costs nothing.
But Charnas draws a hard line between AI that helps him and AI that acts for him. He has AI draft and reply to emails in his voice, but he approves every one before it goes out. When he’s looking for tenants for an off-market space, he’ll ask AI to name the 10 likeliest prospects and search his email and Dropbox for existing connections. Then he decides what happens next.
“It comes up with a great idea, but it’s not reaching out on its own,” he said.
The real question, Charnas said, isn’t whether to use AI tools but what you let them touch.
“An AI agent that can only help you write is pretty low risk,” he said. “If it’s handling wire instructions, client data, building access, [tenant] screening, that’s a different level of control.”
Accountability doesn’t transfer to the software, he added. If an AI agent makes a mistake on behalf of a real estate professional, the human real estate professional will get the blame.
“It drafts, it summarizes, it can show a vision, but it doesn’t actually make the commitment,” Charnas said. “‘The tool went rogue’ or ‘the process was sloppy’: that’s when you get burned.”
He doesn’t think those risks will slow AI adoption in the real estate industry. They’ll just shape it. “Firms are going to set permissions and keep human sign-off,” he said.
‘Slow down fast’
State regulators are taking up the issue, Hawkins said, through ARELLO, the association of state real estate licensing officials.

Dominik Pogorzelski
ARELLO is the trade organization that brings together real estate licensing departments from different states, because they all have different rules. Hawkins said he will soon meet with ARELLO executive leadership, and they are “bringing that conversation straight to AI.”
“They’re all getting together to ask: What should the rules be? What should the enforcement be?” Hawkins said. “Can we issue best policies and practices? Because the regulations, as you know, are way behind what’s happening in AI today.”
Even AI’s longtime industry builders are urging caution. Dominik Pogorzelski, president of MLS at Restb.ai, whose computer vision technology reaches more than 1 million agents through their MLSs, told Hawkins on a recent episode of his Inside AI podcast that the industry needs to “slow down fast.”
“I think it’ll get crazy before it settles down,” Pogorzelski said.
Hawkins doesn’t want the industry to retreat from AI. One client told him losing his AI tools would cost him at least three hours a day. Hawkins fears a high-profile failure could prompt regulators or MLSs to shut tools down wholesale, taking the productivity gains with them.
“Real estate is a highly regulated industry. All of a sudden, someone could say you can’t do this because you’re not meeting their threshold for acceptance,” Hawkins said. “You don’t have the gates and guardrails built in, and you can’t be transparent about how everything works, whether it’s training or data retention. And if there’s a violation, they’ll say, ‘I’m sorry, I have to take that away from you.'”
Hawkins advises real estate agents to use professional tools with built-in safety protections rather than building their own. “Even if you’re a tech person, how much time do you really have?” he said. “Being a real estate agent is a full-time job, and so is being a tech person.”
He said the goal is to keep people accountable at every step that matters.
“We’ve got to understand that we are the human in the middle,” Hawkins said. “We’re the humans who evaluate and review. We can’t lose that with AI. We can’t just let it go off on its own. That’s where the danger is.”