US watchdog outlines 16 AI cybersecurity policy options amid testing gaps
A new GAO assessment identifies four policy goals and 16 options for managing AI-related cyber threats, while warning that testing standards and guidance remain inadequate.
The United States Government Accountability Office has set out 16 policy options for strengthening cybersecurity in an era of increasingly capable artificial intelligence, highlighting weaknesses in testing, governance and information sharing. Its technology assessment, released on November 20, 2026, examines both the ways AI can amplify cyberattacks and the defensive applications that security teams are adopting.
The watchdog organized its options under four goals: addressing gaps in cybersecurity guidance and practices, closing research gaps, governing AI tools throughout their lifecycle and improving workforce capabilities. The options are proposals for consideration by policymakers, standards bodies and other institutions, not binding requirements or new federal mandates.
GAO said generative AI can help malicious actors create convincing phishing material and deepfakes, while agentic systems may automate multiple stages of attacks. Its report cites previously documented cases of AI-assisted malicious activity, including efforts to test malware against security products and attacks that exploited stolen credentials and weak authentication. Those examples were drawn from industry reporting and should not be interpreted as new incidents uncovered by the assessment itself.
The assessment also describes benefits for defenders, including analyzing network traffic, examining source code for vulnerabilities and processing large volumes of security information. Yet it cautions that widely accepted criteria for measuring offensive and defensive AI capabilities are lacking. Experts consulted by GAO raised concerns about privacy, data quality, integration costs, fragmented regulation and the difficulty of validating AI-generated outputs.
Among the policy options are expanding testing environments, requiring human authorization for consequential AI-initiated actions, evaluating AI products before acquisition, validating security-tool outputs and maintaining fallback capabilities for critical functions. The report does not assert that any single option has been adopted or that it would eliminate cybersecurity risks.
The report follows broader public warnings about AI-enabled attacks, but its immediate significance is institutional: it offers lawmakers and agencies a documented set of choices for procurement, testing and oversight. Independent coverage of the release also emphasized the four policy goals and 16 options, while distinguishing these recommendations from enforceable rules.
Spotted an error? Request a correction · Our editorial principles · Contact the newsroom