Sakana AI has announced the launch of its Security Research Group, an independent R&D team studying collective, AI-driven cyber defense as well as the security and safety of AI systems themselves. According to the [company's announcement](/security-research-group/), the group operates separately from Sakana AI's commercial business and is tasked with serving the public interest by publishing research, reproducible benchmarks, and practical guidelines, and by collaborating with government, industry, and academia.

The group frames its approach around the idea that intelligence and resilience emerge from collectives, described in the announcement as a school of fish rather than a single swimmer. Sakana AI states that as a builder of frontier AI it shares responsibility for the security of the society that uses such systems.

Why it matters

The announcement cites a sharp rise in disclosed software flaws. NIST reports that CVE submissions rose 263% between 2020 and 2025, and that early 2026 ran nearly a third above the same period a year earlier. JVN iPedia, which provides vulnerability information in Japanese, registered nearly twice as many entries in the latest quarter as a year earlier, most of them sourced from the NVD.

AI is widely suspected to play a part, though the announcement notes that more confirmation is needed. FIRST points to AI-assisted discovery alongside administrative factors, and NIST staff say AI-based discovery tools have added to their workload. Japan's government now treats the trend as a national security matter, and the strain is visible: NIST has stopped routine analysis of older records and prioritizes flaws known to be exploited, while Japan's financial regulators have asked institutions to treat the response as a management issue with top management directly involved.

Sakana AI summarizes the imbalance as flaws being found at machine speed and fixed at organizational speed, with the people in between bearing the cost. Attackers are also using AI agents to run intrusions against many targets at once, and AI systems are themselves becoming targets: NICT recorded more probing aimed at AI service endpoints in the first quarter of 2026.

Technical details

The group argues that faster triage and patching are necessary but insufficient. It cites CISA's review of two fiscal years finding that many attackers exploit well-known, straightforward weaknesses, with input validation and memory safety flaws the most dependable ways in, and calls for eliminating persistent, preventable weaknesses. NICT's alert on attacks against the servers behind mobile apps names similar weaknesses, such as missing authorization checks and unvalidated input. The FBI's account of AI-agent intrusions lists SSRF, IDOR, missing API authentication, and injection.

The proposed answer is upstream: preventing flaws at the design and build stage rather than chasing them after release. 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. The group describes four layers of defense, from upstream to downstream: the quality of each product, the design and configuration of the system, operation and patching, and detection and response. It works across all four, with long-term weight on the upstream two.

The announcement invokes the Kano model from quality management, which sorts quality into attractive, one-dimensional, and must-be categories. Safety and security are presented as must-be qualities: unnoticed when present, strongly dissatisfying when missing, and therefore requiring treatment as quality requirements from the start rather than inspection at the end.

What to watch

Sakana AI positions quality as an area where Japan can lead, citing the country's manufacturing reputation, and proposes turning the same AI capability that finds flaws in finished systems into a way of checking quality at every step of design, building, and operation. The group also suggests that building the expertise of human reviewers into AI agents could extend quality assurance in scale and depth, and that flaws found today could inform how future systems are built. Concrete outputs, including the promised reproducible benchmarks and guidelines, have not yet been detailed.

About the author

Sakana AI (JP)

Thoai Dispatch Curated — curated source featured on Thoai Dispatch.

More by Sakana AI (JP)