Lean SuperIntelligence raises pre-seed for on-prem, self-improving security AI
What's the deal? Palo Alto-based Lean SuperIntelligence (LSI)Dealroom has a profile for this one. Try Dealroom → has closed a pre-seed round backed by Fusion FundDealroom has a profile for this one. Try Dealroom →, alongside angels who have led AI and security at OpenAI, CrowdStrike, SentinelOne, Zscaler, Commvault, and Gruve AI. The startup builds self-improving cybersecurity large language models (LLMs) that run on customers' own infrastructure.
What's the endgame? LSI pairs two models that train each other: LSI-Offense-1 finds, exploits, and validates vulnerabilities, while LSI-Defense-1 detects, investigates, and responds. Inside what the company calls Cyber World Models, every attack the offense finds becomes a lesson for the defense, and every defense that holds forces the offense to find a new path.
The models are customizable and air-gapped, lean enough to scan every code check-in continuously.
The results: On independent offense benchmarks, LSI-Offense-1 outperforms frontier models including Claude Mythos™, GPT-6, and Gemini-3.8-Cyber at autonomously discovering and exploiting vulnerabilities, and has already uncovered previously unknown flaws. On public defense benchmarks, LSI-Defense-1 outperforms rival models at a fraction of the cost, returning verdicts in seconds with fewer false positives.
Why now? Attacks move in minutes — a rumoured vulnerability can become a working exploit in under 10 minutes. Enterprises increasingly want to own their security, not rent it: two-thirds moved AI workloads back on-prem last year, and sovereign cloud spending is set to hit $80B in 2026.
Who's behind it? Founder and chief executive officer Chandra Khatri holds more than 50 patents and publications. He was founding head of AI at Krutrim, where he built India's first sovereign LLM in 2023, and co-founded Got It AI, building the first enterprise LLM and a hallucination guardrails model that was acquired. He previously led AI work at Uber AI, Amazon Alexa AI, and eBay Research.
"You cannot defend against an attack you have never seen. So we trained the Offense LLM first," Khatri said. "Frontier labs give everyone the same model. We provide each enterprise its own customizable model, running on-premise and trained on its own private data."
The signal: LSI is betting that specialised, lean models can beat general-purpose frontier systems where it counts — proof that in security, enterprises may favour ownership and precision over scale.
Read more: prnewswire.com
Image credit: U.S. Army Combat Capabilities Development Command
