Lean SuperIntelligence, a Palo Alto-based AI lab, said it has closed an initial financing round and brought out two self-improving cybersecurity models that run on customer premises. LSI-Offense-1 finds, exploits and validates vulnerabilities, while LSI-Defense-1 detects, investigates and responds, the company said. The two models are trained against each other, so an attack found by the offense model informs the defense model, and a defense that holds pushes the offense model to look for a new route, with both improving on the customer's own data.
On independent, real-world offense security benchmarks, the company said LSI-Offense-1 outperforms frontier models including Claude Mythos, GPT-6 and Gemini-3.8-Cyber at autonomously discovering, exploiting and validating vulnerabilities, and that it has already uncovered previously unknown flaws. On public defense benchmarks spanning multi-step threat detection, investigation and response, LSI-Defense-1 outperforms models from the same labs at a fraction of the cost, returning verdicts in seconds with fewer false positives, according to the company.
The models are customizable and can be air-gapped, and LSI said they are lean enough to run continuously, scanning every check-in software developers make. The pre-seed round is backed by investors and angels who have led AI and security work at OpenAI, CrowdStrike, SentinelOne, Zscaler, Commvault and Gruve AI, the company said. LSI was founded by Chandra Khatri, who was founding head of AI at Krutrim and held AI roles at Amazon Alexa, Uber and eBay Research.
LSI-Offense-1 and LSI-Defense-1 are available now to a limited number of design partners, among them security vendors, managed security service providers and certain Fortune 500 enterprises, for air-gapped, on-prem, sovereign or managed deployment. Access can be requested at lsi.inc.
