The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

Sakana AI Launches Security Research Group in Tokyo

By The Daily Byte ·21 Nov 2026 ·4 min read
Tokyo Tower, the red lattice tower in Minato City, Tokyo, seen from below

Tokyo-based Sakana AI has launched an independent Security Research Group, a research team that will study collective, AI-driven cyber defence and the security and safety of AI itself. The group was announced on 20 November 2026 and will operate separately from the company's commercial business, with a public-interest mission to publish research, reproducible benchmarks and practical guidelines, and to work with government, industry and academia.

The group is led by Akihiro Umegai, Sakana AI's Chief of Security Research, and David Ha, the company's co-founder and chief executive. It is hiring researchers in Tokyo for its technical staff. Its starting point is an idea that runs through Sakana AI's work: intelligence and resilience emerge from collectives, like a school of fish rather than a single swimmer.

Why now: flaws found at machine speed

The announcement sets out a picture of defenders falling behind. The US National Institute of Standards and Technology reports that submissions for CVE identifiers, the standard labels for disclosed software flaws, rose 263 percent between 2020 and 2025, and that early 2026 ran nearly a third above the same period a year earlier. Japan's JVN iPedia vulnerability database registered nearly twice as many entries in its latest quarter as a year earlier. NIST has stopped routine analysis of older records and now prioritises flaws known to be exploited.

Attackers are changing too. The announcement notes that attackers are using AI agents to run intrusions against many targets at once, and that AI systems are themselves becoming targets: Japan's National Institute of Information and Communications Technology recorded more probing aimed at AI service endpoints in the first quarter of 2026. Flaws, in Sakana AI's phrase, are being found at machine speed and fixed at organisational speed, and the people in between are bearing the cost.

Prevention instead of patching

The group's argument is that faster triage and patching, while necessary, will not bring overall risk down on their own. A review by the US Cybersecurity and Infrastructure Security Agency found that many attackers exploit well-known, straightforward weaknesses, with input validation and memory safety flaws the most dependable ways in. The FBI's account of AI-agent intrusions lists the same kinds of weakness: server-side request forgery, insecure direct object references, missing API authentication and injection.

The lasting answer, in the group's view, lies upstream, in how systems are designed and built, so whole classes of flaws are made unreachable before a product ships. Japan's government-wide package points the same way, asking software vendors to apply secure-by-design principles and to use advanced AI across development to find and fix vulnerabilities early. Sakana AI frames security as a "must-be" quality in the Kano model of quality management: unnoticed when present, and a cause of strong dissatisfaction when missing, which is why it is easily left out of the definition of quality altogether.

Three lines of research

The group sets out three research areas. Prevention by design asks how the qualities a system must keep can be defined before it is built, how automated fixes can be checked against criteria the automation itself cannot rewrite, and how a bad fix can be undone quickly. Collective defence with AI agents asks whether defensive agents distributed across organisations can stay in sync, so that an attack seen once becomes a protection shared by all who take part. Open questions include what may be shared and with whose consent, and who is accountable when agents act on a shared signal.

The third area is securing AI itself, starting with the defenders' own agents. Agents that triage and patch hold privileged access, so the first two areas depend on this one: how to limit each agent's authority to its task, trace why it acted, and keep tools built for defence bound to defence.

What it builds on

The group builds on Sakana AI's existing work, including Fugu-Cyber, its cyber model released on 21 July 2026, which coordinates specialised models and which Sakana AI reports at 86.9 percent on the CyberGym benchmark and 72.1 percent on CTI-REALM. Those are the company's own figures, described by Sakana AI as comparable to leading cyber-focused frontier models, and the announcement itself cautions that benchmark results do not necessarily carry over to real enterprise environments. Measuring that gap is one of the group's stated research questions.

It will operate under three guidelines: vendor neutrality, with its own systems evaluated under the same transparent, reproducible conditions as any outside tool; responsible disclosure, sharing white papers, benchmarks and defensive playbooks openly while withholding actionable exploit material; and building the technical bridge between Japan's security frameworks and day-to-day operational practice. As Umegai put it in the announcement, "security and safety are not a brake on AI's progress but the condition for sustaining it."