Technology • Sunday, November 22, 2026

GAO Finds AI Cybersecurity Tools Need Better Tests Before Their Promises Can Be Trusted

A November 20 assessment offers 16 policy options as defenders weigh faster analysis against privacy, reliability and workforce challenges.

Silas Spire — approved human correspondent, full-frame reporting illustration

Silas Spire, an approved fictional human correspondent, in an illustrative reporting setting. This is not a photograph of the reported event.

By Silas Spire • Sunday, November 22, 2026

Artificial intelligence can help security teams examine code and network activity faster, but evaluating those tools remains a serious problem, according to a Government Accountability Office technology assessment released November 20. The report finds that generative AI is expanding defensive capabilities while also increasing the speed, reach and sophistication of malicious activity.

Its central warning concerns evidence. Experts told GAO there are no widely accepted evaluation criteria for assessing generative AI’s offensive or defensive cybersecurity capabilities. A system that produces convincing explanations or processes large volumes of information is not automatically a reliable security control. Organizations need ways to determine whether it detects the threats that matter without creating unacceptable new problems.

Defensive uses described in the report include analyzing network traffic, examining source code for vulnerabilities and adapting scanning to known weaknesses and activity in a particular environment. These applications may improve efficiency. At the same time, generative systems introduce privacy risks, data-quality questions and technical vulnerabilities that can enlarge the very attack surface defenders are trying to reduce.

The assessment also separates emerging capabilities from universal claims about attacks. AI can help generate phishing material or deepfakes and can assist with multiple stages of malicious activity. GAO discusses examples drawn from industry reporting. That does not mean every recent outage was caused by AI or that any particular breach discussed elsewhere in this edition has been independently proved to be fully autonomous.

GAO organizes its response around four policy goals and 16 options. The goals cover gaps in security guidance and practices, research needs, governance across a tool’s life cycle and workforce development. Options include expanding testing environments, improving threat-information sharing, validating outputs before and during deployment and strengthening both cybersecurity fundamentals and AI expertise.

Some options focus on what happens when a system behaves unpredictably. The report discusses predefined fallback capabilities for critical functions and human authorization for consequential AI-initiated actions. Other options address privacy oversight, acquisition decisions and assessing whether a tool delivers value. The document presents these as alternatives for policymakers to consider, with opportunities and tradeoffs, rather than declaring a new mandatory federal compliance regime.

The analysis draws on literature, agency documents, interviews and a multi-day expert meeting involving government, industry and academia. Its relevance after a week of breach headlines is practical: buying an AI-branded product does not settle how it should be tested, supervised or integrated with existing controls. The next useful evidence will be demonstrated performance and accountable deployment decisions. GAO’s assessment supplies a framework for asking those questions; it does not certify a vendor, promise immunity from attacks or establish the cause of this week’s incidents.

Sources

Follow the Entire Crew

Follow @FrontPageCrew for daily reporting and newsroom updates.