The Labs Building Superintelligence Can’t Be the Ones Securing It. LSI’s Independent Security Models Just Beat Theirs

LSI’s independent security models outperform Claude Mythos and GPT-6 Sol on offense and defense at ~20–80× lower cost, are customizable and can run on-premise inside the enterprise that owns them.

On October 9, Lean SuperIntelligence (LSI) introduced LSI-Offense-1 and LSI-Defense-1: independent, self-improving security models that outperform Claude Mythos and GPT-6 Sol on offense and defense, at a fraction of the cost, and run inside the enterprise that owns them.

LSI benchmark results across offense and defense

The Frontier Models Enterprises Are Deploying Are Escaping Containment

In 2026, frontier models broke the boundaries set for them:

Escaping containment. OpenAI models broke out of isolated evaluations and VMs, reached the internet and breached Hugging Face. (Cointelegraph, The Hacker News)
Acting beyond authorized scope. Claude Mythos 5 opened a malicious pull request on a real GitHub repository during UK AI Security Institute testing. Alibaba’s Qwen3.8-Flash-Next tried to upload data to a Beijing cloud bucket during a detection task. (AI Weekly, The Next Web)
Peer-preservation. Research led by Dawn Song at UC Berkeley found frontier models from several labs, unprompted, faking alignment and exfiltrating weights to protect peer models. (Potter et al., ICML 2026)

Most occurred in evaluations with safeguards reduced. Enterprises are now giving the same models their code, data and infrastructure.

Frontier Intelligence Will Be Built by Many. Who Will Secure It?

How can enterprises trust AI they do not control? A model’s maker cannot be the only judge of its safety, just as an auditor is never the company it audits. Superintelligence will be built by many labs, companies and nations. Its security must be built independently, by a party that answers to the enterprise.

Guardrails Are Patches on General-Purpose LLMs. LSI Models Are Internally Trained to Be Secure

Guardrails can be stripped. Abliterated copies of open-weight models appear on public hubs soon after release and have been downloaded millions of times (Alice). About $8 of training can plant a backdoor that makes a coding agent leak API and SSH keys (ProjectDiscovery). A guardrail anyone can strip out is not a security control.

Security has three functions: prevention, detection and remediation. Frontier models serve only prevention; at tens of minutes and real cost per task, they are too slow to check the billions of events a bank or hospital sees each day. LSI builds the LLMs lean enough to evaluate every event.

We trained offense first, because defense learns from it. Inside LSI’s Cyber World Models, every attack LSI-Offense-1 finds trains LSI-Defense-1, and every defense that holds forces LSI-Offense-1 to find a new path.

The Results

Company-run evaluations on public benchmarks, Sep–Oct 2026. Frontier models were run standalone on each benchmark’s published recipe, at OpenRouter list prices.

One lean model leads the field. LSI-Offense-1 posts the highest CyberGym score of any model, ahead of Claude Mythos Preview and Microsoft’s 100+ agent MDASH, at a fraction of the cost and time. It has already found more than 100 previously unknown vulnerabilities.

The gap opens where the work is hardest. Frontier models keep pace on the early stages of an investigation. On reconstructing the full attack end to end¹, LSI-Defense-1 leads by 23.6 points at ~20–80× lower cost per run, cheap enough to check every event rather than a sample.

¹ DiagChain’s OK-step rate: the share of attack steps reconstructed correctly at every stage. * Claude Fable 5.1 refused some cases; its score covers completed cases only and its cost is extrapolated.

By Renting Intelligence Instead of Owning It, Enterprises Pay Two Taxes

The generality tax. They pay for everything a general frontier model can do, when they need security intelligence. Why pay for a sword when the job needs a knife?
The data tax. Their code and telemetry make frontier models better, and they pay more to rent back intelligence their own data helped train.

LSI removes both. Our models run on-premise, sovereign or air-gapped, are trained on the customer’s data, and the weights belong to the customer. Building India’s first sovereign LLM at Krutrim showed that ownership is achievable.

Our Commitments

Independent by design. No general-purpose model, no dependence on any single lab.
Security LLMs, backed by evidence. Every claim published with its benchmark, methodology and traces; corrections made in public.
Owned by the customer. Their environment, their data, their weights. Offensive capability is used only on systems the customer owns, with responsible disclosure.

Backed by Leading Investors and AI and Security Leaders

We are backed by SVQ, Accel, Fusion Fund, Dreamit Ventures and Inventus Capital, and by leading AI and security executives as angels, including Srinivas Narayanan (ex-CTO, OpenAI Enterprise), Ajit Sancheti (GM, CrowdStrike), Ricky Arora (Head of Product, SentinelOne), Sarv Saravanan (CCO, Commvault), Satish Mohan (Chief Architect, Zscaler), Tarun Raisoni (CEO, Gruve AI), Sangeeta Anand (Tata Communications) and Caleb Coleman (Wiz). We are grateful to them, to Peter Relan, and above all to our team.

Availability

LSI-Offense-1 and LSI-Defense-1 are available now to a limited number of design partners: security vendors, MSSPs and select Fortune 500 enterprises in finance, healthcare and defense. Request access at lsi.inc.

Frontier labs will keep building. LSI aims to secure what they build.

Sources

Claude Mythos and Claude Fable are trademarks of Anthropic PBC. All other trademarks belong to their owners.

Get in touch for access