"Defence only" is a business model now

When the general labs ship offensive capability behind a vetting queue, the differentiated product is not access. It is incentives you can put in a contract.

Corma raised $60 million led by Sequoia, with Khosla and Coatue, to be a frontier lab dedicated to defensive security.

Sixty million buys a layer, not a model

That number is a large seed and a small foundation-model budget, which locates the bet precisely. Corma is not training the capability; it is building what turns a capability into a defender — context about a specific network and the judgement to act without breaking production.

Those are the scarce inputs in security. Raw capability stopped being scarce the week a general lab shipped a model completing 95 percent of advanced offensive tasks.

Concentration makes specialists possible

The same quarter that saw MGX close $49 billion and two labs absorb 43 percent of all venture dollars also funded a $60 million specialist and a $6.5 million robotics seed. Those are not contradictory. Extreme concentration at the model layer pushes returns toward the application layer, because that is the only place left where a small team can own something.

The line will be tested

Defence and offence use the same techniques; the distinction lives in authorisation and intent, both of which sit in customer contracts rather than in the model. Plenty of security companies have started here and found the line moved when a large enough buyer asked. And the buyer will ask — the capability is already commercially available to anyone who passes vetting.

A company whose entire pitch is that it will not do the other thing has one asset: never having done it.

AI Weekly — AI News Today, August 10 — Top AI Stories & Live Updates → · Crunchbase News — The Week's 10 Biggest Funding Rounds →