Skip to main content
Back

Sakana AI Opens a Public-Interest Lab to Fight AI-Powered Cyberattacks

Sakana AI spins up an independent cyber defense lab in Tokyo, betting that quality engineering upstream beats patch-chasing downstream.

  • Sakana AI launches the Security Research Group, an independent R&D team run separately from its commercial business
  • Co-led by Chief of Security Research Akihiro Umegai and CEO David Ha, based in Tokyo
  • Two themes: AI driven collective cyber defense, and the security and safety of AI systems themselves
  • Context: CVE submissions up 263% from 2020 to 2025, with NIST deprioritizing older records
  • Builds on prior Sakana work including Digital Red Queen, ShinkaEvolve, and Fugu-Cyber (86.9% CyberGym)
  • Hiring senior researchers in Tokyo, with open publication of papers, benchmarks, and playbooks

Tokyo-based Sakana AI has launched the Security Research Group, an R&D unit focused on two connected problems: using AI to support collective cyber defense across organizations, and protecting AI systems from attacks, misuse, and unsafe behavior. Chief of Security Research Akihiro Umegai and CEO David Ha will co-lead the group, which operates separately from the company’s commercial business and plans to publish its work openly.

The unit’s public-interest charter includes collaboration with government, industry, and academia. Its hiring notice also acknowledges the work’s dual-use potential, since methods for finding flaws can help defenders and attackers. The group says it will withhold details that could enable exploitation and use coordinated disclosure, giving affected maintainers private notice and time to fix vulnerabilities before publication.

Vulnerability volume outruns review

The Common Vulnerabilities and Exposures system assigns CVE identifiers to reported software flaws, while the US National Institute of Standards and Technology adds severity and affected-product data through its National Vulnerability Database. NIST figures cited by Sakana show that CVE submissions rose 263% between 2020 and 2025, with early 2026 running nearly a third above the same period a year earlier. To manage the backlog, NIST has stopped routinely enriching older records and now prioritizes vulnerabilities known to be exploited.

Attackers are also automating several stages of an intrusion through tools such as ARTEX, an open-source framework in which multiple agents powered by large language models coordinate their actions. Sakana cites an FBI FLASH alert reporting that threat actors have adapted ARTEX to run attacks with limited manual intervention. Its entry points include server-side request forgery, which tricks a server into making unintended requests; insecure direct object references, an access-control failure; missing API authentication; and injection flaws. Automation can increase exploitation volume using these established bug classes without depending on novel zero-days.

Patch queues cannot solve design flaws

Sakana expects vulnerability growth to outrun the capacity of teams responsible for triage and patching. A US Cybersecurity and Infrastructure Security Agency review covering two fiscal years found that attackers repeatedly exploited well-known, straightforward weaknesses, with input-validation and memory-safety flaws providing especially reliable routes into systems. Eliminating recurring defect classes before release would reduce the number of findings that reach patch queues.

Sakana treats security as a software-quality property and uses the Kano model from manufacturing to explain its approach. Under that model, “must-be quality” describes a baseline that customers assume will be present and whose absence can cause severe failure. The group applies that logic by placing security requirements into architecture, implementation, testing, and procurement from the beginning. It also argues that Japan’s manufacturing culture offers familiar methods for designing quality into production processes.

Contracts, legacy systems, and thin teams

  • Outsourced operations: Vendors often maintain production systems, so patch ownership and response duties must be defined in contracts.
  • Technical debt: Legacy stacks and incomplete asset inventories leave teams unsure which products are deployed and which vulnerabilities apply.
  • Limited staff: Security and IT teams cannot expand quickly, requiring organizations to concentrate scarce expertise on their highest-risk systems and decisions.

Japanese financial regulators have asked institutions to handle cyber response as an executive management issue, with direct involvement from senior leaders. Because staffing cannot expand quickly, Sakana is focusing automation on repeatable work while reserving human expertise for higher-risk decisions.

Three bets for AI-assisted defense

  1. Prevention by design: Use AI during architecture, coding, and testing to remove recurring vulnerability classes before deployment.
  2. Collective defense: Coordinate agents across organizations so they can share observations and support a joint response.
  3. Secure defense agents: Harden the agents themselves because their privileged access makes them high-value targets.

From Core War to enterprise defense

  • Digital Red Queen: In Core War, programs compete for control of simulated computer memory. Sakana evolved programs written by large language models against their predecessors, and later generations transferred better to unseen human-written opponents. The research question is whether that adversarial co-evolution can be adapted safely to real red-team and blue-team work, meaning offensive testing and active defense.
  • ShinkaEvolve: This method evolves code using relatively few evaluations. Sakana sees potential applications in generating detection rules and candidate patches, where testing against real systems can be expensive.
  • Fugu-Cyber: This system coordinates specialized models and scored 86.9% on CyberGym and 72.1% on CTI-REALM, two cybersecurity evaluations. Sakana reports that the results are comparable to leading cyber-focused frontier models, while cautioning that benchmark performance may not carry over to enterprise environments. Measuring that transfer gap is part of the research agenda.

The first deliverables: papers and playbooks

Developers and security teams can expect papers, reproducible benchmarks, practical guidelines, and defensive playbooks as the group’s initial outputs. Sakana has not announced a product, API, or release schedule, and its warning about benchmark transfer means published scores will need validation against real codebases, infrastructure, and operating constraints.

If the collective-defense research succeeds, agents could exchange observations, correlate activity, and coordinate responses across companies or public bodies. Such a system would move some detection beyond isolated security operations centers, or SOCs, into a shared network. Its implementation would require clear rules for data access, privacy, trust, and accountability across organizational boundaries.

Sakana is hiring senior researchers in Tokyo and requires conversational Japanese and business English. The role emphasizes problem formulation, critical thinking, incident write-ups, failure analyses, and published research, indicating that the group wants researchers with practical security experience as well as academic credentials.