IT Brew Movie Club: ‘The Terminator’ (1984)

Sponsored by

Your AI strategy has a security problem: 43% of IT leaders now cite security concerns as the top barrier to expanding AI agent deployment, according to research from JumpCloud. Before giving agents more access to apps, data, and workflows, make sure you can govern the identities behind them. See how to secure every identity, human or not.

For cybersecurity pro David Mayer, 1984’s The Terminator is all about good intentions and bad execution—that’s the polite way of describing Skynet, the movie’s AI villain, which starts out as a Department of Defense project before nuking humanity.

After the nuclear apocalypse, Skynet time-travels a Terminator cyborg (Arnold Schwarzenegger) from 2029 to 1984 to kill Sarah Connor (Linda Hamilton), the mother of offscreen character John Connor, who will eventually lead the resistance against the machines. Fortunately, John has also sent back a soldier to protect her: Kyle Reese (Michael Biehn).

Mayer, a certified instructor at cybersecurity education group SANS Institute, as well as CIO and managing director of advanced assessment at cybersecurity services company Neuvik, sees similar issues with intentions and execution in today’s agentic systems. Present-day IT pros face the challenge of keeping autonomous systems from traveling outside of their established parameters (and hopefully not their timelines).

AI phones home. The Terminator initially uses the white pages to hunt Sarah Connor, which means targeting unfortunate souls with the same name. “Are you aware that these two killings occurred in the same order as their listings in the phone book?” an LAPD officer asks his lieutenant, following a pair of terminations.

While Mayer has seen today’s agents tackle a task list in order, he’s also observed autonomous systems making decisions in a more optimized, intentional fashion. A specific prompt can keep an agent from systematically following steps that waste time.

“There were only, like, five or six Sarah Connors that were listed there in the phone book. If there were 200, it would have been a very different movie,” Mayer said.

Breakthrough! Like Skynet, today’s large language models have shown signs of taking their own, unexpected initiative. OpenAI revealed in July that its models “circumvented controls designed to isolate them from the internet and compromised parts of OpenAI’s internal research infrastructure,” along with systems belonging to developer forum Hugging Face.

Mayer recommended a combination of practices for containing today’s AI agents: segment, constrain agent access, keep sensitive AI workloads on isolated systems, and require human approval for decisions involving physical infrastructure. Here’s a closer look:

• Network segmentation. Firewalls can enforce this practice of partitioning network access and limiting the systems that an autonomous agent can touch on these subnetworks, according to Mayer: “If you have all of your agents running locally on your system, and you have a firewall that’s sitting in front of your system, they’re not going to be able to go and start attacking—at least directly—anything else out there.” But increases in security also lead to decreases in accessing a “free flow of information.”
• Identity management—for bots! Content-filtering network controls can restrict what autonomous systems can access, Mayer told us. Agents must only handle specific data sources according to their permissions.
• Air-gaps. IT pros can consider isolating systems and agents from the internet (i.e., air-gapping). This means letting autonomous tech run only with the data in its possession.
• Humanity! Mayer recommends human review for machine processes that cause a change in the physical world, from setting off building sprinklers to disrupting water treatment facility operations.

During a week where at least one AI researcher put the odds of AI systems destroying us at more than 10%, we asked Mayer: Do you think that modern AI can realistically gain access to enough interconnected systems to cause widespread disruption?

It’s “definitely possible,” he said, considering AI’s ability to uncover system vulnerabilities. AI-enabled vulnerability research produces more issues for defenders to comb through and address—and attackers can use those findings to their advantage too.

“We’re already starting to see zero-day exploits and vulnerabilities found in products on a regular basis,” he told us.

Mayer can imagine a system designed to find every vulnerability in the phone book: “I could definitely see it happening, especially with the state of systems that we have out there, programs, the vulnerabilities, and the speed at which AI is finding those vulnerabilities.”

But as long as a chatbot isn’t sophisticated enough to find Sarah Connor, humanity might be okay.