AI Files: Why Humans Are the Biggest AI Security Threat

Sophisticated algorithms matter less than the physical fragility of the people managing them. Protecting the future of intelligence requires looking past the code to secure the humans behind it.

Transcript

SPEAKER_02 0:00

Eloise, a recent essay makes a powerful and somewhat unsettling claim right at the outset. AI is not dangerous. People are. That flips our common anxieties about artificial intelligence on their head. What's the core idea behind this argument?

SPEAKER_01 0:17

It's a crucial distinction, Noah. The essay contends that AI systems themselves do not sabotage their own operations. The real risk and the actual source of vulnerability comes from human actions. We are talking about everything from insider revenge, espionage, and public protests, to targeted strikes during wartime on the physical data centers and supercomputing hubs that are absolutely essential for AI to function.

SPEAKER_02 0:44

So it is not about Skynet becoming self-aware, but rather about a disgrantled employee or a nation state taking down the servers. Let's look at some of the examples cited in the essay, starting with what it calls peacetime insider sabotage. These incidents suggest human actions are already causing significant damage. Can you walk us through the case of the Virginia Twin Brothers?

SPEAKER_01 1:11

Certainly. In December 2025, Munib and Sohaib Akdar, former federal contractors at OPEX, were arrested. They used lingering access credentials after their termination to delete approximately 96 government databases. These databases contained everything from Freedom of Information Act records to Department of Homeland Security investigative files, and IRS data with personally identifiable information for over 450 people. The damage was immense. Public records were wiped, agency services were disrupted, and sensitive data was stolen, posing a national security risk.

SPEAKER_02 1:49

That is an astounding breach, disrupting government functions and compromising sensitive data. How are they caught? And what does this specific case teach us about protecting AI infrastructure?

SPEAKER_01 2:01

A multi-agency investigation involving agencies like the FDIC and DHS OIGs plus HSI caught them after spotting the deletions and critically an AI query they made about clearing logs. They had even discussed wiping their devices. The lesson here is clear. Immediate access revocation upon termination and robust log monitoring are essential. These measures can prevent ex-insiders from causing such extensive damage.

SPEAKER_02 2:33

That makes perfect sense. Beyond sabotage, the essay also highlights intellectual property theft. The case of Ling Wei Ding, an ex-Google engineer, sounds particularly impactful.

SPEAKER_01 2:45

It was indeed. In January 2026, Ling Wei Ding was convicted for stealing over 2,000 pages of Google's AI supercomputing secrets. These included designs for TPU and GPU chips, orchestration software, and SmartNICs used in AI data center workloads. He was secretly working for Chinese firms at the same time. The damage was enormous. He essentially handed blueprints for rivaling US AI infrastructure to an adversary, potentially costing billions in a competitive edge.

SPEAKER_02 3:17

This is a direct transfer of advanced AI capabilities. How can organizations guard against such sophisticated theft?

SPEAKER_01 3:26

The FBI caught Ding by monitoring data uploads to his personal cloud and investigating his foreign affiliations. The lesson is simple but critical. Companies must meticulously track employee outside work or side gigs and monitor for attempted data exfiltration. Insider IP theft is the fastest way to arm adversaries. And the essay underscores that this applies directly to AI development.

SPEAKER_02 3:53

Moving from individual actions to collective ones, the essay also discusses peacetime protest sabotage. The Nepal CG Net Satungal Data Center Arson is a stark example.

SPEAKER_00 4:06

Yeah, in September 2025, protesters set fire to the CG Net Satangal data center amidst broader unrest. This resulted in a nationwide internet blackout for hours, even days in some areas. Servers were destroyed, though services were partially restored later. It was tied to a wave of vandalism during those protests. This incident shows that public anger over infrastructure projects can escalate to low-tech arson, causing significant disruption.

SPEAKER_02 4:34

So it is not necessarily high-tech cyber warfare, but fundamental physical destruction. The essay also points to rising tensions around data centers in the U.S. What should we learn from these protest-related incidents?

SPEAKER_01 4:48

The events in Nepal and even disruptive protests in places like Port Washington, Wisconsin, and Claremore, Oklahoma, highlight that opposition to new AI campuses or data centers can turn volatile. While the U.S. incidents did not involve physical damage, the Sofan Center's November 2025 report warned that these tensions, often fueled by concerns over jobs or energy consumption, could escalate to violence. The key lesson here is the importance of early and extensive community outreach before construction. Preventing small sparks from becoming large fires is cheap insurance. Indeed, in March 2026, Shahed drones directly hit two AWS facilities in the UAE, with a nearby strike damaging a site in Bahrain. This caused structural damage, power loss, and water damage from fire suppression systems. The result was localized outages for various services like EC2 and S3, and customers were advised to migrate traffic. Iran even confirmed responsibility, claiming the facilities had links to the U.S. military.

SPEAKER_02 5:58

This illustrates a profound shift in conflict where digital infrastructure becomes a physical battleground. What is the takeaway for securing our increasingly AI-dependent world?

SPEAKER_01 6:09

This case makes it abundantly clear that geopolitical risk is a primary concern for data centers. They need what the SA calls missile-level redundancy, not just advanced cyber defenses. The SUFON Center further flagged exploding online calls for physical AI data center sabotage in 2025. Ultimately, the SA's core message is that if we secure the humans through better access controls, robust monitoring, effective protest intelligence, and strategic wartime sighting of facilities, the perceived danger of AI itself diminishes significantly.