Self-improving cybersecurity large language models can conduct offensive and defensive operations in customizable local deployment environments, and operate at a significantly lower cost.
Palo Alto, California, USA, October 9th / PRNewswire / -- Lean SuperIntelligence ( LSI ) indicates that the company is the first AI laboratory to create self-improving attack-defense security large language models, which can help enterprises stay ahead of attackers without handing over their data to the cloud. The two models will improve upon each other: LSI-Offense-1 LLM is used for discovering, exploiting, and verifying vulnerabilities, while LSI-Defense-1 LLM is used for detection, investigation, and response within minutes rather than days or weeks. In the LSI's Cyber World Models, each attack discovered by the offensive model becomes an opportunity for the defensive model to learn, and each successful defense forces the offensive model to find new approaches. As a result, both models continue to improve on the customer's own data. These models are customizable and can run in isolation; they outperform leading large models in vulnerability discovery and threat detection, and are light enough to run continuously, scanning every code change submitted by software developers. LSI states that its pre-seed round of financing has received support from several leading investors and angel investors who have previously been involved in OpenAI, CrowdStrike, SentinelOne, Zscaler, Commvault, and AI related to security work.
Outperforms state-of-the-art models on both offensive and defensive fronts.
In independent real-world attack-defense security benchmark tests, LSI-Offense-1 and LLM have significantly outperformed leading models, including Claude Mythos™, GPT-6, and Gemini-3.8-Cyber, in terms of autonomously discovering, exploiting, and verifying vulnerabilities. These models have identified vulnerabilities that were previously unknown. The results show that specialized, lightweight network security models can surpass general-purpose leading models.
In public defense benchmark tests that cover multi-step threat detection, investigation, and response, LSI-Defense-1 and LLM significantly outperformed similar laboratory models at a much lower cost. They were capable of diverting alerts, correlating telemetry data, making judgments within seconds, and had fewer false positives.
The significance lies in
- Attacks are measured in minutes. A rumored vulnerability can potentially turn into exploitable code in less than 10 minutes; defenses must keep up with the pace with each incident.
- Companies prefer to own rather than rent. LSI indicates that nowadays, companies send their code and telemetry data to cutting-edge laboratories, spending hundreds of millions of dollars on inference, yet they end up owning nothing in return. The company argues that this is also why two-thirds of enterprises migrated their AI workloads back to local systems last year,[1] and why sovereign cloud spending is expected to reach 80 billion dollars by 2026.[2]
LSI provides enterprises with a self-improving security model that they can possess.
The company was co-founded by the team that developed national/sovereign large language models, as well as the first batch of AI Guardrails, Alexa, Uber AI, and Salesforce Research AI laboratories, in collaboration with top security experts.
- LSI Founder and CEO. Chandra Khatri holds over 50 patents and papers. He served as the AI founding leader of Krutrim where he developed India's first sovereign large language model in 2023. He is also a co-founder of Got It AI, where he created ELMAR – the first enterprise-level large language model, as well as TruthChecker – the world's first hallucination protection model (later acquired). He has also led projects at Uber AI, Amazon Alexa AI, and eBay Research AI.
- The founding engineers of LSI have built a sovereign AI and foundational models from pre-trained data to production deployment, serving tens of millions of users in the process.
- The founding security team of LSI has discovered hundreds of critical CVE vulnerabilities for large institutions and national intelligence agencies.
Comment
"You can't defend against an attack that you've never seen before. So we focused on training our offensive capabilities first," said Chandra Khatri, the founder and CEO of LSI. "Our LSI-Offense-1 and LLM are designed to be a generation ahead of attacks in the real world, while our LSI-Defense-1 and LLM learn from every attack discovered by LSI-Offense-1. The cutting-edge laboratory provides everyone with the same model. We, on the other hand, offer each company its own exclusive, customizable model that runs locally and is trained using their private data."
"Chandra's career has always progressed with every major wave in AI: the enterprise large language models of Alexa, Uber AI, Got It AI, as well as guardrails; and also India's first sovereign large language model. LSI is the natural next step." —__ Founder and Managing Partner of Fusion Fund Lu Zhang
Availability
LSI-Offense-1, LSI-Defense-1, and LLM are now available to a select group of design partners, including security vendors, MSSP, as well as some Fortune enterprises with over 500 employees. They can be used for isolated, local, sovereign, or managed deployments. Access requests can be submitted through lsi.inc.
About LSI
Lean SuperIntelligence ( LSI ) is the first laboratory to build a self-improving large-scale language model for offensive and defensive security, with the goal of advancing towards Security SuperIntelligence. LSI was founded by Chandra Khatri ( Krutrim , Got It AI , Amazon Alexa , Uber AI ), bringing together creators of sovereign and enterprise-level large language models as well as senior security researchers.
Follow LSI: LinkedIn, X/ Twitter
Claude Mythos is a trademark of Anthropic PBC. All other trademarks are the property of their respective owners.
Reference materials
- 66% of enterprises have migrated their AI workloads back from public clouds, according to a survey of 1,500 enterprise architects conducted in June 2026.
- Global sovereign cloud IaaS expenditures are expected to reach $80 billion in 2026, according to data from February 2026.
Media Contact
SOURCE LSI Inc.
