- The Washington Times - Friday, November 20, 2026

Artificial intelligence is helping hackers move faster and strike harder, while the people defending against them still lack widely accepted ways to test whether their AI tools work, according to a new government watchdog report.

As a string of outages hit Microsoft, Apple, JPMorgan Chase and others this month, a Government Accountability Office report released Friday found that autonomous AI agents have begun carrying out cyberattacks with minimal human direction, and there are no widely accepted criteria for evaluating the AI tools meant to stop them.

“AI improves the success rate of each stage of attack,” GAO said.

People were interested in these podcasts

AI lets hackers make phishing and deepfake content that is more convincing and more precise, the report said. It also speeds up every stage of an attack, shrinking reconnaissance “from weeks to hours or less.”

Shown is a Chase Bank location in Media, Pa., Monday, Nov. 16, 2026. (AP Photo/Matt Rourke)
Shown is a Chase Bank location in Media, Pa., Monday, Nov. 16, 2026. (AP Photo/Matt Rourke) Shown is a Chase Bank location … more >

Agentic AI is an emerging trend, with autonomous agents carrying out multiple attack stages with minimal human input.

The FBI’s Internet Crime Complaint Center reported more than $20 billion in losses in 2025 across over 1 million complaints, including more than $893 million tied to AI-related cybercrime.

AI-leveraged attacks are hard to count because no definitive technical indicators separate them from traditional attacks, GAO said.

Defenders use AI for threat detection, incident response, malware reverse-engineering, penetration testing and vulnerability discovery. One survey cited by the GAO found that over 70% of analyst tasks can be automated.

But there are no widely accepted ways to evaluate AI security tools, and defenders often rely on vendor claims. Experts said benchmarks “become useless almost immediately.”

Generative AI is nondeterministic, so identical inputs can yield different outputs. That makes audit trails and validation hard.

A computer screen shows the Microsoft support page in London, Friday, Nov. 13, 2026. (AP Photo/Kirsty Wigglesworth)
A computer screen shows the Microsoft support page in London, Friday, Nov. 13, 2026. (AP Photo/Kirsty Wigglesworth) A computer screen shows the Microsoft … more >

Agents that mix trusted instructions with untrusted data are vulnerable to prompt injection and “agent hijacking,” the GAO said.

Other obstacles include fragmented regulation — 80 of 117 federal cyber regulations the GAO reviewed carry reporting requirements — outdated information-sharing law, slow procurement processes and unpredictable token-based costs.

A skills gap compounds the problem, and experts also said automation is eroding the entry-level work that trains analysts.

The GAO said the technology is dual-use and the speed advantage favors attackers.

Contact the author

Copyright © 2026 The Washington Times, LLC. Click here for reprint permission.

Please read our comment policy before commenting.