Lilith.
⌕
Editorial illustration: Sakana AI bets Japan’s cyber defense on shared agents, not faster patching
Lilith illustration · editorial remix

Sakana AI has created the Security Research Group, a research team separated from its commercial business. Akihiro Umegai and David Ha lead the Tokyo group. It plans to publish papers, reproducible benchmarks and practical guidance while working with government, industry and academia.

Defense is meant to learn across organizations at machine speed

The program has two main themes: collective AI driven cyber defense and the security of AI systems themselves. Sakana wants to explore a network of defensive agents where an attack observed by one organization improves protection for all participants. It explicitly leaves consent, signal trust and accountability for agent actions as open questions.

The second layer is prevention. The team wants to define security properties before a system is built, verify them continuously and check automated fixes against criteria the agent cannot rewrite.

Japan’s constraint is how systems are built, not only attack speed

Sakana grounds the program in rising vulnerability disclosures. Cited NIST data says CVE submissions increased 263% from 2020 to 2025. The authors also acknowledge that AI’s contribution cannot yet be cleanly separated from administrative factors.

The practical focus is Japanese organizations with outsourced operations, technical debt and limited specialist staff. Faster discovery merely expands the queue if teams keep building the same preventable weaknesses. The group therefore shifts attention from patching toward secure by design practice and quality during development.

Benchmarks still describe a laboratory, not critical infrastructure defense

Sakana builds on Fugu-Cyber, which the company reports scored 86.9% on CyberGym and 72.1% on CTI-REALM. The announcement itself says those results may not transfer to real enterprise environments and names that gap as a research question.

Independence currently remains an organizational promise from the company funding the work and supplying its starting technologies. The group promises to publish evaluation conditions, compute budgets and harnesses. Its credibility will depend on testing Sakana systems as harshly as outside tools.

The first open tests will separate an institution from a manifesto

Reproducible benchmarks, clear responsible disclosure rules and partnerships that provide operational evidence without leaking sensitive data will matter most. Collective defense also needs firm limits on what agents may share and who can stop a faulty automated response.

The team is still hiring senior researchers. Once it publishes methods, baselines and failed experiments, readers can judge whether public interest truly takes priority over the company narrative.

Lilith's verdict

Sakana wants to turn isolated guards into a school that passes warnings before an attack reaches the next neighbor. Without clear command, the whole school can follow the first bad signal just as quickly.