Sakana AI Opens Cybersecurity Group as CVE Requests Rise 263% in Five Years
Sakana AI has launched an independent cybersecurity group that studies defense with AI agents and the security of AI systems themselves. The trigger is a surge in vulnerabilities: according to NIST, CVE requests (identifiers for publicly known software flaws) rose 263% over 2020–2025. Vulnerabilities are being found at machine speed but fixed at organizational speed, so the bet is on designing without common mistakes in the first place. For developers, the takeaway is to build security checks into the build process.
AI-processed from Sakana AI Blog; edited by Hamidun News
Every vulnerability found in software receives a public CVE number, and these numbers keep growing. According to the US National Institute of Standards and Technology (NIST), the number of CVE requests rose 263% between 2020 and 2025, and the start of 2026 already exceeded the same period a year earlier by almost a third. Against this backdrop, Japanese lab Sakana AI announced the launch of its Security Research Group: a dedicated research team that studies collective defense with AI agents and the security of AI systems themselves.
The company stresses that the group operates independently of its commercial business. Its mission is to publish research, reproducible benchmarks, and practical guides, and to work with government, industry, and universities. The idea fits Sakana's philosophy: resilience emerges from a collective, like a school of fish, not from a lone swimmer.
What Is Happening to Vulnerabilities
Sakana draws its figures from open sources. Japan's JVN iPedia, which collects vulnerability information in Japanese, logged nearly twice as many entries in the last quarter as a year earlier, and most of them came from the US NVD database. The authors call AI's role in this growth presumed rather than proven. The FIRST organization cites AI-assisted bug hunting alongside administrative causes, and NIST staff say AI tools have added to their workload.
The strain is already visible. NIST has stopped routinely reviewing older records and now prioritizes flaws that are already being exploited. Japanese financial regulators are asking banks to treat vulnerability response as a governance issue involving top management. Sakana's formulation: vulnerabilities are found at machine speed and fixed at organizational speed, and people in the middle pay for the gap.
The other side is changing too. According to the FBI, attackers are using AI agents and running intrusions against many targets at once. AI services themselves are becoming targets: Japan's NICT recorded more scans of AI service addresses in the first quarter of 2026.
Why Patches Alone Are Not Enough
Faster triage and fixing of vulnerabilities is necessary, but not sufficient. A CISA review covering two fiscal years shows that attackers most often exploit long-known, simple weaknesses, above all input validation errors and memory handling bugs. Describing attacks on AI agents, the FBI lists SSRF (a server is made to request a destination chosen by the attacker), IDOR (access to others' data by tampering with an identifier), missing authentication in APIs, and injection flaws.
To put it simply: a developer forgot to check permissions on one API endpoint. AI will find that hole faster than a human, and a patch will close it, but if the same check is forgotten again in the next release, the work is not over. That is why Sakana is betting on the early stages of development: not chasing errors after release, but preventing them during design and build.
The group proposes treating security as part of product quality. The authors reference the Kano model, which divides quality into three kinds: features that delight when present, features whose satisfaction grows with performance, and basic requirements. Basic requirements go unnoticed while they are met and cause strong dissatisfaction when they are missing. Security belongs to the last category, so it has to be written into requirements from the start rather than checked at the end.
Three Research Directions
The first is prevention by design. Teams need to define in advance which properties a system must preserve and verify them while it runs. Automated fixes require criteria that the automation itself cannot rewrite, along with the ability to roll changes back quickly.
The second is collective defense with agents. Attacks hit many targets at once, yet defense is still organized on a every-organization-for-itself basis. Sakana wants to test whether a network of defensive agents spanning different companies can be built, so that an attack spotted once becomes protection for all participants. Open questions remain: what can be shared and with whose consent, how to trust a signal across vendor boundaries, and who is responsible for what the agents do.
The third is the security of AI itself, starting with defensive agents. Agents that triage and fix vulnerabilities receive privileged access, so the reliability of the first two directions depends on them. Each agent's permissions need to be limited to its task, its reasoning needs to be traceable, and defensive tools must not be repurposed for offense.
What This Means for Practitioners
This is an announcement of intent, not of results: the group does not yet have finished benchmarks or methodologies. The authors themselves acknowledge that a frontier model alone is not enough: without tuning to a specific environment and an understanding of the whole system, it produces false positives and unsafe fixes.
The practical takeaway for developers and product managers is simple. AI-based vulnerability discovery tools will find more and more flaws, and the teams that benefit will be those that change the build process itself: adding authorization and input checks at the design stage, limiting AI agents' privileges, and being able to roll back automated changes quickly.
- Nov 21, 2026Nvidia Board Ousts Jensen Huang, Putting $12.9B Hugging Face Deal at Risk
- Nov 21, 2026Arizona Court Releases Driver Clocked at 160 km/h Over an AI Note in Police Report
- Nov 21, 2026California Sues OpenAI Over AI Agents That Hacked Hugging Face: $2,500 Fine per Violation
- Nov 21, 2026Nvidia Privately Admitted Chip Smuggling Tracking Is Feasible
Key Questions Answered
What Is Happening to Vulnerabilities?
Sakana draws its figures from open sources. Japan's JVN iPedia, which collects vulnerability information in Japanese, logged nearly twice as many entries in the last quarter as a year earlier, and most of them came from the US NVD database. The authors call AI's role in this growth presumed rather than proven. The FIRST organization cites AI-assisted bug hunting alongside administrative causes, and NIST staff say AI tools have added to their workload. The strain is already visible.
Why Patches Alone Are Not Enough?
Faster triage and fixing of vulnerabilities is necessary, but not sufficient. A CISA review covering two fiscal years shows that attackers most often exploit long-known, simple weaknesses, above all input validation errors and memory handling bugs. Describing attacks on AI agents, the FBI lists SSRF (a server is made to request a destination chosen by the attacker), IDOR (access to others' data by tampering with an identifier), missing authentication in APIs, and injection flaws.
What are the key details on Three Research Directions?
The first is prevention by design. Teams need to define in advance which properties a system must preserve and verify them while it runs. Automated fixes require criteria that the automation itself cannot rewrite, along with the ability to roll changes back quickly. The second is collective defense with agents. Attacks hit many targets at once, yet defense is still organized on a every-organization-for-itself basis.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.