AI in Cybersecurity Gets 16 GAO Policy Options

AI in Cybersecurity Gets 16 GAO Policy Options

Estimated reading time: 7 minutes · Last updated:

AI in cybersecurity cuts both ways, the Government Accountability Office (GAO) concluded in an assessment released on 20 November 2026. Generative AI lets attackers produce convincing phishing lures and deepfakes and strike with greater speed, precision and accuracy. The same technology helps defenders scan networks and find flaws in source code. To steer adoption, GAO set out four policy goals and 16 policy options for legislators, agencies, standards bodies and industry. They range from larger AI testing environments to human sign-off before AI takes consequential actions. The work, catalogued as GAO-27-108458, drew on literature reviews, stakeholder interviews and a multi-day expert meeting. GAO presents the options as a menu for policymakers, not a set of recommendations, and says the list is not exhaustive.

Key takeaways

  • Two-sided technology: GAO's assessment GAO-27-108458, released on 20 November 2026, finds AI both amplifies cyberattacks and expands what defenders can do.
  • Agentic attacks: In a June 2026 case cited by GAO, several AI agents tested malicious code against different security products to refine detection evasion.
  • Basics still fail: A low-skill attacker in February 2026 used commercial generative AI services and stolen credentials against devices lacking multi-factor authentication.
  • Policy menu: GAO lists 16 policy options under four goals, with the governance goal alone carrying seven of them.
  • Measurement gap: Experts consulted by GAO said no widely accepted criteria exist for judging the offensive or defensive capability of generative AI.

How Agentic AI Lets Attackers Run Whole Campaigns

GAO says AI gives threat actors a way to stretch the reach and impact of familiar attacks. Polished phishing content and deepfakes become cheap to produce, and attacks land with more precision, which raises both their speed and their scale. The trend GAO singles out is agentic AI handling several stages of an attack with little human involvement. Agentic AI is a system with enough autonomy to make its own decisions and adapt as conditions change.

One case from June 2026 shows what that looks like. A threat actor used AI to build a testing lab for malware that evades detection, and investigators found multiple AI agents checking malicious code against different security products. The agents mined security research for new techniques, matched them to known frameworks, ran them in test environments and reported back.

A second case, from February 2026, is cruder. A low-skill attacker combined several commercial generative AI services to mount a widespread campaign, reportedly using credentials stolen in earlier breaches against devices with weak passwords and no multi-factor authentication. GAO drew both incidents from industry reporting. The February case matters because AI amplified ordinary neglect rather than exotic skill.

What Generative AI Adds to Network Defense

Defenders are not starting from zero. GAO notes that various forms of AI have supported cybersecurity tools for decades, and that recent generative AI advances, the category that includes ChatGPT, have widened what defenders can do. Teams use machine learning and generative AI to make their tools, staff and operations more efficient.

Network traffic analysis is the clearest example GAO offers. Adding generative AI to these tools enables dynamic scanning that takes account of newly disclosed vulnerabilities and of activity inside the environment being watched. Defenders can also point generative AI at source code to spot weaknesses, and use it to sift large volumes of network data quickly.

GAO argues these gains leave defenders better placed to handle a rising volume of malicious activity. The scale of that activity is the backdrop to the whole assessment: GAO says state and non-state attackers hit government and private-sector systems thousands of times daily, threatening water, energy, telecommunications and financial services. It also says the government's pace still lags the threat.

Why Defenders Struggle to Trust Their AI Tools

Most of the risks raised by the experts GAO consulted concern generative AI specifically. The headline problem is measurement: there are no widely accepted criteria for evaluating what generative AI can do on offence or defence. A security team buying such a tool therefore has no common yardstick for comparing it with rivals or with the tools it already runs.

Generative AI also brings fresh risks of its own that can enlarge the attack surface, the total set of points an intruder could exploit. Experts sorted the remaining obstacles into four groups: data privacy and quality; technical issues rooted in how generative AI is built; governance and trust problems caused by fast-moving technology and outdated frameworks; and organizational barriers such as cost, an AI skills gap and difficult integration.

Two problems drew repeated mention. Multiple experts named fragmented regulatory approaches and insufficient data sharing as key challenges. Both are coordination failures rather than technical ones, which is why several of GAO's policy options target standards and information flows, not the models themselves.

Four Goals and 16 Options for Policymakers

GAO built its policy goals from executive orders, national strategies, its own High Risk List, federal laws and other sources. The four goals are closing gaps in cybersecurity guidance and practice, filling AI-related research gaps, governing AI tools across their cybersecurity lifecycle, and strengthening the cybersecurity workforce. Policymakers here means more than Congress: GAO includes agencies, standards organizations, industry and other groups.

The guidance goal carries four options: harmonizing cybersecurity requirements, better sharing of threat information, closing AI-related privacy gaps and applying security fundamentals more consistently. The research goal has three: broader research into safe AI adoption, innovation in identity management for AI agents, and larger AI testing environments.

The workforce goal has two options. One keeps defenders' current cybersecurity skills from eroding; the other builds AI expertise among them. Pairing the two is a deliberate hedge, given the February 2026 case showing that weak passwords and missing multi-factor authentication still give attackers easy entry.

Fallbacks and Human Sign-Off for High-Stakes AI

Governance is the largest goal, with seven options, and the most technical. It rests on a distinction GAO draws between two kinds of AI. Deterministic AI follows strict rules and gives predictable outputs. Nondeterministic AI, including generative models, involves probability or randomness, so the same input can produce different answers.

That distinction drives two options. One would give critical cybersecurity functions a deterministic fallback, meaning a predefined alternative such as rule-based tools or manual procedures that keeps the function running if the AI fails, goes offline or behaves outside its intended use. Another would limit autonomous and nondeterministic AI in consequential roles until it is certified, though GAO does not describe who would certify it.

The remaining options would make a person approve consequential actions an AI initiates, apply a risk-based approach to buying AI products, check AI tool outputs before and during deployment, test whether AI solutions deliver real value, and create a senior role responsible for AI privacy risk.

GAO's four policy goals for AI in cybersecurity
Policy goal Options Example option Page in GAO-27-108458
Guidance and practices gaps 4 Harmonize cybersecurity requirements 41
AI-related research gaps 3 Expand AI testing environments 43
Lifecycle governance of AI tools 7 Human approval of consequential AI actions 45
Cybersecurity workforce 2 Strengthen AI expertise 49

Can policy keep defenders ahead of AI-powered attackers?

The case for

  • Generative AI already lets defenders scan code for vulnerabilities and process large volumes of network data quickly, according to GAO.
  • Larger AI testing environments and agreed evaluation criteria would give buyers a way to compare tools they currently cannot measure.
  • Deterministic fallbacks and human sign-off would keep critical functions running if an AI system fails or misbehaves.

The case against

  • Experts told GAO there are no widely accepted criteria for judging generative AI's offensive or defensive capability.
  • Fragmented regulation and thin data sharing were named by multiple experts as key obstacles, and neither has a quick fix.
  • GAO says the government still does not move at a pace matching the threat, while agentic AI lets attackers automate several attack stages.

What to be careful about

  • Defenders may deploy nondeterministic AI in critical functions without a fallback, leaving those functions exposed if the system fails.
  • Generative AI added to security tooling can itself widen the attack surface it is meant to protect.
  • Organizations chasing AI capability may neglect basics such as strong passwords and multi-factor authentication, the gap exploited in the February 2026 case.

The bottom line

GAO's assessment treats AI in cybersecurity as an arms race in which both sides draw from the same toolbox. The incidents it cites point in two directions. Agentic attacks that automate research, testing and reporting are a genuinely new threat. Credential attacks on devices without multi-factor authentication are an old one made cheaper. The 16 options reflect that split, pairing research and testing with workforce basics and governance safeguards. Their weak point is measurement: without agreed criteria for judging generative AI, buyers cannot easily tell which tools help. GAO has offered a menu rather than a mandate. Whether agencies, standards bodies or Congress pick anything from it is the open question.

What to watch

  • Watch whether any congressional committee or agency takes up GAO's proposal for human approval of consequential AI-initiated cybersecurity actions; no date has been set.
  • Watch for standards organizations to begin work on shared evaluation criteria for generative AI security tools, the gap experts flagged to GAO; no timetable has been announced.
  • Watch whether a certification scheme emerges for autonomous and nondeterministic AI in consequential functions, which GAO's restriction option presupposes; nobody has proposed one or set a date.

Frequently asked questions

What is agentic AI and why does GAO worry about it in cyberattacks?

Agentic AI is a system with enough autonomy to make decisions and adapt its behaviour on its own. GAO flags it because attackers can use it to run several stages of an attack with minimal human input, as in a June 2026 case where multiple AI agents tested malware against different security products.

Are GAO's 16 AI cybersecurity policy options binding?

No. GAO presents the 16 options as a menu of alternatives that policymakers could consider, and says the list of goals and options is not exhaustive. Policymakers in this sense include legislative bodies, agencies, standards organizations and industry.

Who can answer questions about GAO-27-108458?

GAO lists three contacts: Kevin Walsh, director of Information Technology and Cybersecurity; Karen L. Howard, director of Science, Technology Assessment, and Analytics; and Jennifer R. Franks, acting chief technology officer. Media inquiries go to Sarah Kaczmarek, managing director of the Office of Public Affairs.



Share:

Categories

Newest course every month

Advertise your offline course to a wider audience with our landing page.

You May Also Like

GAO finds AI in cybersecurity magnifies attacks and strengthens defense, and lists 16 policy options across four goals to help...
US tech outages prompted Senators Cantwell and Rounds to press AI czar Jay Clayton on AI-found flaws, as officials say...
Microsoft MAI Code 2 nearly matches rival coding models at a fraction of the cost, Microsoft says, alongside a new...