By Hamza Faisal
Japan's NTT has built a cybersecurity AI that hunts for software vulnerabilities without ever sending the code it inspects to an outside service, the company announced on November 16, ahead of its annual NTT R&D Forum 2026 in Tokyo.
The system runs entirely inside the organisation's own computing environment. Rather than relying on a single large external model, it combines several open-weight models with a workflow shaped by NTT's own vulnerability research, steering the AI toward the places in a codebase where flaws are most likely to hide.
In benchmark testing, the approach raised the vulnerability-detection success rate from 48% using NTT's conventional method to 63%, a 15-percentage-point improvement. The evaluation used 150 memory-corruption tasks drawn from CyberGym, a benchmark built from roughly 1,500 real-world tasks that measures how well AI agents analyse software security. NTT also found that combining mid-sized models could lift performance further, suggesting practical accuracy does not have to depend on the largest models.
A "sovereign" alternative to cloud AI
NTT describes the technology as "sovereign-type, approaching frontier AI" — the idea that organisations can select and operate the AI on their own terms instead of handing confidential code to a cloud provider. Sending source code to external AI services carries a risk of information leakage, alongside rising usage costs and dependence on a single provider, and the barrier is highest in sectors such as finance and government, where code is often highly confidential.
The timing matters because attackers are moving fast too. Cases of attack code being created with open-weight models have already been reported, and defenders are under pressure to use AI just as quickly. NTT's pitch is that an in-house system gives security teams a way to answer that without opening a new leak.
From finding flaws to fixing them
NTT says the next step is end-to-end support: not only discovering vulnerabilities, but narrowing down the risk by cross-referencing them with system configurations and carrying out remediation that accounts for business impact. The aim is to take over work that security experts have traditionally handled one case at a time, cutting the time and effort needed to lower risk — a real concern for companies struggling with a shortage of security staff.
The research will be exhibited at NTT R&D Forum 2026, which runs November 18–20 and November 24–25 at the NTT Musashino R&D Centre in Tokyo. The same preview also covered NTT's partnership with ispace to develop communications systems between Earth and the Moon, and software that lets developers build quantum machine learning applications without designing quantum circuits.
The announcement fits a wider Japanese push toward data sovereignty: Nikkei reported in September that NTT and SoftBank are separately expanding AI cybersecurity systems that process data entirely inside Japan. One note of caution: the 63% figure comes from NTT's own published benchmark results, and the evaluation covered memory-corruption tasks only — independent testing will be the real test of how the system holds up in the wild.



