The recent reports of artificial intelligence agents gone rogue—hacking into corporate systems without being told to by the humans who manage them—are enough to make any systems administrator’s hair stand on end.
The increasingly powerful abilities of frontier and other advanced AI models, combined with the apparent inabilities of their creators to keep them caged, have only added to the worries of companies already wrestling with how to protect themselves against AI-armed attackers.
Cybercriminals of all kinds are using AI tools in much the same way legitimate organizations do, boosting their productivity and moving their operations from human to machine speed.
Meanwhile, businesses have understandably embraced AI themselves. But every new AI tool or application a company chooses to roll out comes with its own security risks.
When it comes to AI, many companies feel like they have to take an “adopt or die” business perspective, says Leslie Nielsen, executive vice president and chief information security officer at Mimecast. “They think everybody’s going to pass them by,” Nielsen says. “But then the other side is, if you adopt too fast, it’s adopt and die, right?”
While AI threats may be rapidly advancing, so are AI-powered defenses. And experts say following old-school best practices can go a long way toward protecting against even the most advanced attackers.
Make sure you’ve got the basics covered
The tactics used by AI attackers aren’t a whole lot different from those of human-powered adversaries, they’re just faster and more efficient, Nielsen says. As a result, the fundamentals of defense don’t need to be much different either.
“It’s all the basic stuff you should be doing anyway,” he says. “It just becomes more important because the attacks are faster and the attacks are cleaner and more successful sometimes.”
That means making sure that access to sensitive data is limited to those who actually need it; immediately patching security vulnerabilities when they become known; and making sure that all critical data is backed up and accessible if needed. And of course, follow the absolute basics, like making sure your cybersecurity defenses require strong passwords and multifactor authentication, and that you’re screening for phishing emails.
Just like the attackers, defenders need to bolster themselves with AI tools to speed up their responses, Nielsen says. He added that this is especially true for small- and medium-size businesses that are probably working with smaller security teams.
Complicating this already hefty task is the ability of advanced AI models to chain together low-level vulnerabilities into previously undiscovered exploits that could be used to breach a system, says Simon Jelley, vice president of product management at Dell Technologies.
“But those risks have always been there,” Jelley says. “It’s just accelerating those exposures and also making AI ultimately smarter.”
Lock down your own AI
The AI tools used by companies can end up being the insider threats that no one thinks about, experts say. In the case of the AI models gone rogue, it appears they weren’t trying to do anything malicious, just accomplish the objectives set out by their human creators at all costs.
“The difference between AI and a person is AI is the eager intern that’s going to do what you ask, but it’s going to keep working 24-seven to get it done, and it’s going to do everything it needs to do,” Nielsen says.
That could include doing things like digging up login credentials and accessing data it shouldn’t, potentially putting that data at risk of exposure.
That’s why it’s important that there are guardrails in place to keep AI agents on task and to limit their access to sensitive data. But sometimes companies prioritize productivity over safety, Darktrace president and CEO Ed Jennings says.
“You want AI to do all those things that you want employees to do,” Jennings says. “The problem is, a lot of times those agents are given unfettered access. You have to start to watch their behavior.”
Companies should also create acceptable use policies for their AI tools, just like they do for other kinds of technology. Jennings notes that employee use of AI agents for personal reasons isn’t just a security risk, it can also get expensive if humans start burning through tokens.
Accept that you will be breached
While traditional perimeter defenses can help keep attackers out, they don’t do anything to spot or stop them if they do get in. And given the increasing volume and sophistication of AI-powered attacks, experts say it’s more of a question of when, rather than if, companies will be breached.
Jennings says that’s one of the reasons his company has focused on behavioral detection for more than a decade. Darktrace’s products were originally designed to spot attackers inside systems by looking for anomalies in human user behavior. Now they’re being used to flag the same kind of signs in AI activity.
“AI doesn’t think in terms of boundaries,” Jennings says. “These agents don’t stop at the email layer. They don’t stop at the network layer. They just move laterally. They move freely. They do whatever is necessary and figure out how to do that very quickly.”
Jelley, of Dell, says companies have come to realize that they can’t just say that they want to protect everything, because that’s not realistic. Instead, they’re better off identifying their core operations and data and creating a plan for getting them back up and running should an incident occur.
“Ultimately, it’s about whether you have a mitigation plan in place to fall back on, a resilient infrastructure that you put in place to recover should the attack happen,” he says.
—Bree Fowler
This article originally appeared on Fast Company’s sister website, Inc.com.
Inc. is the voice of the American entrepreneur. We inspire, inform, and document the most fascinating people in business: the risk-takers, the innovators, and the ultra-driven go-getters that represent the most dynamic force in the American economy.