GAO finds no widely accepted yardstick for generative AI’s cyber capabilities
A new assessment pairs evidence of AI-assisted attacks with options for safer defense, including human approval for consequential actions and backup tools.
Loading page…
A new assessment pairs evidence of AI-assisted attacks with options for safer defense, including human approval for consequential actions and backup tools.
Listen to this story
GAO’s November 20, 2026 assessment identifies a measurement and governance gap: there are no widely accepted criteria for evaluating generative AI’s offensive or defensive cyber capabilities. It offers 16 policy options—not new requirements—for lawmakers, agencies, standards bodies and industry, including human approval for consequential actions, ongoing validation and fallback procedures. The practical implication is that adopting AI for cyber operations also requires ways to test its value, limit risky autonomy and keep protection working when AI fails.
In a June 2026 case cited by GAO, multiple AI agents tested malicious code against security products and refined ways to evade detection.
A February 2026 incident involved a low-skill attacker, commercial AI services and stolen credentials; weak passwords and absent multifactor authentication enabled compromise.
Experts cited privacy and data-quality risks, integration costs, skills shortages and fragmented regulation as barriers to defensive AI adoption.
Generative AI gives cyber defenders more ways to find vulnerabilities and sift through network data. But there are no widely accepted criteria for judging its offensive or defensive capabilities, the U.S. Government Accountability Office finds in a report released November 20, 2026. The assessment sets out 16 policy options for putting those tools on firmer footing.
The assessment examines both sides of the same technology: how attackers use AI to increase their reach, and how defenders use it to protect information systems. GAO reviewed literature and agency documents, interviewed government, academic and industry stakeholders, and convened a multiday expert meeting. Its findings distinguish useful defensive applications from the difficulties of testing, integrating and overseeing them.
AI can produce convincing phishing messages and deepfakes while increasing the precision, speed and scale of attacks, GAO says. An emerging development is agentic AI: systems that make decisions and carry out tasks with some autonomy. In cyberattacks, that can mean handling multiple stages with minimal human input rather than simply drafting malicious content.
Drawing on industry reporting, GAO cites a June 2026 case in which an attacker used multiple AI agents to test malicious code against security products. The agents searched security research for attack techniques, mapped them to known cybersecurity frameworks, ran attacks in test environments and reported results. The operation used AI to refine ways of avoiding detection.
Another cited incident, from February 2026, involved a low-skill attacker using commercial generative AI services and stolen credentials. The attacker reportedly compromised devices with weak passwords and no multifactor authentication—an extra login check beyond a password. That case involved gaps in basic security, not just new AI capabilities.
Defenders already use AI to analyze network traffic, identify source-code vulnerabilities and process large volumes of network information. Adding generative AI can make scanning more responsive to newly known vulnerabilities and activity in a particular environment. GAO describes these tools as a way to improve the efficiency of security teams and operations.
Yet generative AI can also create new points of exposure to attack. Experts identified privacy and data-quality problems, technical challenges rooted in model architecture, and governance frameworks struggling to keep pace with AI’s evolution. Cost, missing skills and integration difficulties add organizational barriers. Several experts also highlighted fragmented regulation and insufficient information sharing.
The policy options fall under four goals: closing gaps in cybersecurity guidance and practices, addressing research gaps, governing AI throughout its use in cybersecurity, and strengthening the workforce. Research options include expanding testing environments and investigating how to manage the identities of autonomous AI agents.
Another option is a predictable fallback for critical security functions. GAO describes a predefined alternative, such as rule-based tools or manual procedures, that can keep a function running if AI becomes unavailable, fails or behaves inconsistently with its intended use. This separates continued protection from dependence on a single AI system.
The workforce options pair stronger AI expertise with maintaining existing cybersecurity skills. Alongside better information sharing and basic security practices, they keep conventional defenses within the agenda for adopting more advanced tools.
Loading discussion...
Join the conversation
Explain where you would draw the line between speed and control.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.