AI Agents, Power Demand and Automation Raise US Internet Outage Risk

Key points

• AI agents now sit inside enterprise systems with admin-level tool access, and they give AI-assisted attackers a new way into the edge devices network operators already struggle to patch.
• AI's electricity demand and grid fragility are no longer separate stories: the same data centers that host DNS and cloud control planes can now destabilize the grid that powers them.
• Robot-native logistics means a long Internet outage would stop physical goods from moving, not just websites.
• Taken together, these overlaps make a multi-week outage more plausible and harder to recover from than the network evidence suggests on its own.

US Internet resilience is converging with three fast-moving fields: autonomous AI agents, AI-driven electricity demand and physical automation. Each makes it more likely that failures happen together and harder to recover from them.

Where the lines cross

AI agents and AI cyber offense

On the resilience side, attackers already work at machine speed. Mythos-class models turned up more than 10,000 high or critical vulnerabilities across every major OS and browser in weeks¹, a criminal group used an LLM to build a 2FA-bypass zero-day for a web admin tool², and Chinese state actors hold persistent access in carrier edge routers³. On the agent side, about 40% of enterprise applications are expected to embed task-specific agents by the end of 2026⁴. MCP has become shared tool-calling plumbing at roughly 500M monthly SDK downloads⁵, and one viral agent's skill marketplace was about 12% malware within weeks⁶. One theoretical result argues prompt injection cannot be solved at the filter level⁷.

Together, these mean every agent with credentials to a network management console becomes a door an attacker can reach without touching the router. Cheap inference⁸ and agent task horizons that double every few months⁹ also make it cheap to run long, patient campaigns across many ISPs at once. We think it likely that the first multi-provider outage caused by AI will reach its targets through an agent or tool supply chain, not a direct exploit.

AI power demand and grid concentration

The IEA finds AI-focused data center electricity use rose 50% in 2025, and it attributes about half of US demand growth through 2030 to data centers¹⁰. The resilience evidence shows what that load does to the grid. Data centers have dropped more than 1,000 MW within seconds during disturbances¹¹, DOE models a sharp rise in outage hours as load outpaces firm capacity¹², and a replacement transformer takes 2.5 to 5 years to arrive¹³. The October 2025 us-east-1 outage showed that one DNS fault in Northern Virginia can act as a national failure¹⁴.

What once looked like a spending story is now a stability loop. The demand that finances new capacity also eats the margin that would absorb an attack on the grid. The places where AI load is concentrated (Northern Virginia, Texas) are the same places that host the control planes most Americans depend on.

Physical automation and outage duration

Amazon runs more than 1 million robots, and robots touch 75% of its deliveries, while headcount per facility has fallen to a 16-year low¹⁵. US industrial robot installations rose 11% in 2025¹⁶. Fulfillment designed around robots assumes constant connectivity and power. When staffing is thin, the site cannot fall back to manual work at scale.

This changes what an outage costs. A regional grid or network failure lasting weeks, the timescale that transformer loss¹³ or router reimaging³ implies, would plausibly halt the flow of goods in that region too. The labor evidence points the same way. Entry-level hiring in AI-exposed occupations sits about 19% below trend¹⁷, which over time plausibly thins the pool of people who learned network operations by doing the work. That pool is who you need when routers have to be rebuilt by hand, carrier by carrier.

What the overlap opens up

Neither field shows this alone: the investments that make the country more efficient also shorten the time it can survive disconnected. The risk is that failure modes now travel together. An agent compromise reaches routers, a grid disturbance trips data centers, and a network outage stops warehouses. There is an opportunity too. FERC's order for mandatory reliability standards for computational loads¹¹ and the move toward architectural containment for agents⁷ give operators a point of intervention, if agent permissions, data center ride-through and manual fallback are treated as one resilience problem. The disagreements in the evidence, set out in part 2, mean we hold these convergences with moderate confidence, about 60 to 70%, until they are tested by an actual event.