The U.S. Government Accountability Office said in a report released Friday that artificial intelligence is helping attackers expand the reach of cyberattacks while also giving defenders new tools. The report lays out 16 policy options that policymakers could consider to help manage the technology’s use in cybersecurity.

According to GAO, AI can produce high-quality content for phishing and deepfakes, and it can make attacks more precise and accurate. Phishing is the use of deceptive messages to trick people. GAO said these changes speed up and widen malicious cyber activity. It named a growing trend: attackers using agentic AI, meaning systems that can act on their own, to carry out several stages of an attack with little human input.

The report describes two recent examples drawn from industry reporting. In June 2026, a threat actor used AI to build a testing framework for malware that evades detection. GAO said an investigation found multiple AI agents tested malicious code against different security products. In February 2026, a low-skill attacker used several commercial generative AI services to compromise devices, GAO said. The attacker relied on stolen credentials from past incidents on devices with weak passwords and no multi-factor authentication.

GAO said AI has been used in cybersecurity for decades, but recent advances in generative AI have widened what defenders can do. Defenders use it to make their tools, teams and operations more efficient, the report says. Examples include scanning networks while accounting for newly known vulnerabilities, finding flaws in source code, and sorting through large amounts of network data quickly.

The report also lists challenges. GAO said there are no widely accepted criteria for evaluating the offensive or defensive abilities of generative AI. It said generative AI also brings new risks that can enlarge the attack surface, which is the set of ways a system can be attacked. Experts cited problems with data privacy and quality, technical design, governance and trust, and organizational factors such as cost, an AI skills gap and difficulty integrating the tools. Several experts named fragmented regulation and insufficient data sharing as key obstacles, GAO said.

GAO identified four policy goals through discussions with experts and research. The goals are to close holes in cybersecurity guidance and practices, fill research gaps tied to AI, oversee AI tools across their cybersecurity lifecycle, and strengthen the cybersecurity workforce. The report defines policymakers broadly to include legislative bodies, agencies, standards organizations, industry and other groups.

Among the 16 options are harmonizing cybersecurity requirements, improving threat information sharing and expanding AI testing environments. Others include requiring human authorization for consequential actions started by AI, validating the output of AI security tools, and building fallback capabilities, such as rule-based tools or manual procedures, for critical functions if an AI system fails. GAO said the goals and options are not intended to be exhaustive.

For the study, GAO said it examined published literature and documents from agencies, spoke with stakeholders from agencies, academia and industry, and held a multi-day meeting of experts from government, industry and academia. For the policy goals, it also drew on executive orders, national strategies, its High Risk List, federal laws and other sources.

GAO said in the report that the government is still not operating at a pace that matches changing cybersecurity threats. The full report runs 70 pages. The contacts listed for it include Kevin Walsh, a director in GAO’s Information Technology and Cybersecurity team.