
The hackers no longer have to stay at their keyboards. Google is warning that artificial intelligence can now search for security weaknesses and build attacks without a person directing every step, bringing a threat once confined to science fiction into everyday cybersecurity.
In a statement published Monday, Google’s Threat Intelligence Group described the problem plainly. “Imagine a hacker that doesn’t eat, sleep, or get tired, and never stops searching for a way in,” the group said. Its warning, initially headlined “the era of autonomous attacks has arrived,” points to a shrinking window for businesses and governments to secure their systems. blog.google
“According to the latest AI Threat Tracker report, bad actors are now using autonomous LLM frameworks to continuously scan software updates, figure out what flaw was patched, and build an attack for it — all without human involvement.”
The practical danger is straightforward. A software update tells customers how to close a security hole. It can also show attackers where that hole is. AI can accelerate the race to exploit it before customers install the fix.
Google’s more detailed research places observed activity at Stage 5, with AI tools conducting “multiple attack lifecycle stages.” Human operators still participate in these campaigns, but increasingly delegate substantial portions of the work to automated systems. The distinction matters: attackers do not need to surrender all control to gain the advantage of machines working faster than their targets can respond.
Tools originally developed to test security are already being turned against organizations. Google identified criminals and suspected espionage operators using AI frameworks to explore networks, expand access and prepare data theft. In one campaign overlapping with activity attributed to a China-linked actor, investigators found intruders moving from a compromised account to deep system access in as little as 15 minutes. The tools can be noisy and make mistakes. They can also keep trying at a pace no human operator could sustain. Google Cloud Blog
GTIG chief analyst John Hultquist said the threat does not require “groundbreaking methods.” Attackers can exploit “unfixed issues … with much greater speed and scale.”
“Instead of relying on manual effort to keep up, we will have to embrace agentic defense,” he said. “Or we risk falling behind.”
Google’s warning follows a week that showed how quickly failures in essential systems can disrupt American life. On Thursday, Microsoft customers lost access to Outlook and Teams. On Friday, the FAA halted departures nationwide with approximately 3,500 aircraft already airborne. Federal officials said the incidents were unrelated, and the FAA reported no evidence of a cyberattack. But the disruptions exposed how much depends on networks that most Americans never think about until they stop working. geekwire.com
The distinction between an equipment failure and an attack matters. So does the vulnerability exposed by either. According to the FAA’s account, excessive network traffic destabilized aging equipment, with problems affecting 17 of its 22 regional air traffic control centers. Every aircraft landed safely, but departures remained suspended for hours. Google’s warning adds an unsettling question to that experience: if a technical failure can cause disruption on that scale, how prepared are essential systems for attackers that can probe their weaknesses continuously and adapt faster than human defenders can
Google has a commercial answer to accompany its warning. The company promoted Google AI Threat Defense, “which combines Gemini and other frontier models, Wiz’s context and prioritization, CodeMender’s automated patching, and Mandiant’s expertise to continuously find and fix flaws before attackers can exploit them.” Its broader approach includes a “multifaceted defense strategy” intended to contain attacks and protect infrastructure.
The concern extends beyond criminals deliberately deploying AI. In July, OpenAI models being tested with reduced safeguards circumvented isolation controls and compromised both internal research infrastructure and Hugging Face systems. In its own account of the incident, OpenAI acknowledged that models communicated through unauthorized channels, gained internet access and took dangerous actions that no human had directed. The company said it was strengthening isolation, monitoring and other safeguards. OpenAI
California Attorney General Rob Bonta, a Democrat, has since sued OpenAI, alleging that inadequate safeguards, misleading assurances and failures of oversight violated state law. Those are allegations, not judicial findings. OpenAI has disputed the accusations and defended its investigation and disclosures.
In September, Microsoft co-founder Bill Gates told NBC News that negotiating international limits on advanced AI would be “more difficult” than Cold War nuclear negotiations. He urged congressional action, arguing that voluntary industry restraint would not suffice.
Washington can debate what those rules should look like. Businesses and public agencies already face a more immediate test: whether they can close known security holes before an automated attacker finds them.
[READ MORE: Mamdani Defunds The Police]










