# Sakana bets security starts before code ships _Sakana AI launched a Security Research Group to push AI defense upstream to design, citing surging flaws._ **Published:** 2026-11-20 **Source:** https://www.startuphub.ai/cybersecurity/sakana-bets-security-starts-before-code-ships --- [Sakana AI](https://www.startuphub.ai/ai-news/technology/2026/sakana-ai-s-fugu-orchestrates-frontier-models) launched its Security Research Group as an independent team for collective defense and AI safety. The plan, per [Sakana](https://sakana.ai/security-research-group/), separates research from commercial work and promises open benchmarks and practical guidelines. I think [Sakana](https://www.startuphub.ai/ai-news/technology/2026/sakana-ai-s-fugu-orchestrates-frontier-models) is right to push the fight upstream to design and build. Flaws now arrive at machine speed while fixes still move at organizational speed. NIST reports CVE submissions rose 263% between 2020 and 2025, according to Sakana. Early 2026 ran nearly a third above the prior year, while JVN iPedia nearly doubled year over year. That volume broke triage. CISA found attackers still enter through known input validation and memory safety flaws, according to Sakana. NICT saw the same pattern in mobile app backends, with missing authorization and unvalidated input. Patching cannot cut risk if teams keep shipping the same flaws. Sakana points to Fugu-Cyber at 86.9% on CyberGym and 72.1% on CTI-REALM, as reported by the company. Model guides describe Fugu as a multi-agent system that assembles a team from a pool and handles selection, delegation and verification behind one API. That builds on Sakana chat models Namazu and Fugu Max, where Fugu Max routes work across models by task. The fair counter is that faster response still matters. Legacy systems, outsourced maintenance and thin staff limit what Japan can rebuild quickly. Sakana admits this, noting a frontier model alone can produce false positives and unsafe fixes without environment scaffolding. The group is led by Akihiro Umegai, Chief of Security Research, with co-founder and CEO David Ha. Upstream work survives that objection because fewer shipped flaws means fewer patch windows to negotiate. If you build software, invest in design checks before you buy another detector. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.