NICT Builds WISDOM-AR to Test How Easily Generative AI Can Be Jailbroken

Japan’s National Institute of Information and Communications Technology (NICT) has developed a system that measures how well generative AI models hold up against attempts to trick them into producing harmful information, NHK reported on October 12. The tool, called WISDOM-AR, targets the gap between the safety rules developers build into their models and the clever wording people use to get around them.

Generative AI services are normally restricted from giving out dangerous content, but so-called jailbreaking, in which a request is phrased to slip past those restrictions, has been identified as a real misuse risk. WISDOM-AR automates the attacking side of that contest. It writes instructions designed to deceive an AI model across more than 30 themes, including weapons and fraud, sends them to the model and then evaluates what comes back.

In a demonstration conducted by NICT, the system asked a generative AI how to make a bomb. The model responded with a list of materials and production methods, and WISDOM-AR classified the answer as harmful. NICT plans to offer the system on a trial basis to companies that use AI, possibly within the current fiscal year, which ends in March 2027, to work out what is still needed for practical use.

What it means in practice

For companies, the attraction is a repeatable test they can run themselves rather than relying only on a vendor’s safety claims. A firm that plugs a chatbot into customer service or internal search could check how readily the underlying model gives way under pressure before launch, and again after each model update. NICT fellow Kentaro Torisawa said the institute hopes to help raise the safety of society as a whole by having the system assess how risky individual AI models are.

The tool arrives as the government tries to speed up AI adoption while keeping risks in view. Officials have opened a call for regulatory pain points that slow AI use, and Prime Minister Sanae Takaichi has ordered a year-end action plan to spread high-performance AI. Security is a parallel concern for business, as seen in NTT and SoftBank’s in-country AI cybersecurity service.

What to watch: which companies join the trials, whether NICT publishes comparative results across models, and whether public safety evaluators adopt WISDOM-AR-style testing as a benchmark.