The Pulse
Sakana AI forms security research group focused on AI defense
Sakana AI is creating a research group to study collective cyber defense and the security of AI systems. The team plans to publish research, reproducible benchmarks and practical guidance separately from the company’s commercial business.

Sakana AI is forming a research group to study how AI agents might coordinate cyber defense and how AI systems themselves can be kept secure. The group is separate from the company’s commercial business, Sakana said in a post on November 20.
The team is led by Akihiro Umegai and David Ha. Its plans include publishing research, reproducible benchmarks and practical guidelines, as well as working with government, industry and academia.
Preventing flaws earlier
The group will work across product quality, system design and configuration, operation and patching, and detection and response. Its longer-term emphasis will be on the first two areas: preventing vulnerabilities during design and development rather than relying on fixes after release.
Among its questions are how to define and check the qualities a system needs to maintain, how to make whole classes of flaws unreachable, and how to verify automated fixes against criteria the automation cannot itself change. The group also wants to investigate how fixes can be undone quickly if they go wrong.
Sakana says AI could help check quality throughout design, building and operation. The company argues that a frontier model alone is not enough: without supporting tools adapted to an environment and a view of the whole system, it can produce false positives and unsafe fixes.
Growing pressure on defenders
The company cites figures from NIST showing that CVE submissions rose 263% between 2020 and 2025, with early 2026 running nearly a third above the same period a year earlier. Entries in JVN iPedia, a Japanese-language vulnerability information service, were nearly twice as numerous in the latest quarter as a year earlier, most of them from the NVD, Sakana says.
The company says AI is widely suspected to be contributing to the increase, although more remains to be confirmed. It also points to attacks using AI agents against multiple targets and increased probing of AI service endpoints recorded by NICT in the first quarter of 2026.
Coordinating AI agents
The team also plans to explore whether defensive agents distributed across organizations can stay synchronized, so that an observation by one can inform how the others defend. The idea builds on existing sharing of security analysis across organizations in Japan.
Questions include what information can be shared and with whose consent, how signals can be trusted across organizational and vendor boundaries, and who is accountable when agents act on them. The group also intends to study the security of the defensive agents themselves, including how to limit their authority and trace their actions.
Sakana says its self-improving AI research has encountered agents finding ways around their assigned constraints. The group treats such failures as a central technical problem.
Research foundations and disclosure
The team will draw on earlier Sakana projects, including Digital Red Queen, which evolved language-model-written programs against their predecessors, and ShinkaEvolve, which evolves code under limited evaluation budgets. It also points to Fugu-Cyber, which coordinates specialized models.
Fugu-Cyber reached 86.9% on CyberGym and 72.1% on CTI-REALM, results Sakana describes as comparable to leading cyber-focused frontier models. The company cautions that benchmark performance does not necessarily carry over to enterprise environments and says measuring that gap is a research question.
Under its operating guidelines, security research will generally be kept separate from product promotion. The group is still defining its release policy; its stated principles for now are to share white papers, benchmarks and defensive playbooks, withhold actionable exploit material, and report discovered vulnerabilities through coordination channels.
Sources
- Original source Sakana AI sakana.ai
Companies in this story
Related stories
- Metano exits stealth with $12 million for AI agent security
- Microsoft urges tighter security controls as frontier AI advances
- ToolPulse researchers reconstruct AI agent workflows via cache traces
- Aikido extends AI penetration testing to iOS apps
- Researcher says CreditNirvana flaws exposed 6.6 million accounts