The agent security market arrived before the agent safety science did

A billion-dollar valuation for watching agents inside third-party software, funded in the same week the basic research into multi-agent failure is still taking applications.

Obsidian Security raised $85 million at a $1.1 billion post-money valuation to secure agents operating inside applications the enterprise does not own. Five days earlier, DeepMind and partners opened a $10 million research call on multi-agent safety with applications closing today.

The ratio is the point

Eighty-five million dollars of equity for a product that monitors the problem. Ten million dollars of grants to understand it. That is not a criticism of either — it is a description of how a technology gets deployed when commercial urgency outruns the science.

It has happened before, in exactly this order. Cloud security products shipped years before anyone had a rigorous model of multi-tenant isolation failure. The products were still worth buying. They were just built on operational intuition rather than theory, and the gaps showed up later as incident reports.

What makes agents structurally different

An agent granted access to a SaaS application inherits that application's permission set, and the security team's visibility ends at the tenancy boundary. Monitoring built for humans assumes a session, a login pattern and a rate of action that agents violate on all three counts.

The deeper problem is the one DeepMind's call actually names: agents built by different organisations, with different objectives and no shared operator, negotiating and transacting in the same environment. A system can be individually well-behaved and still produce collusion or cascading failure in a population. No single-model evaluation looks for that.

Where it gets concrete

Naïve raised $28.5 million for an API that lets agents provision infrastructure and stand up companies, reached through Cursor, Claude Code and Codex. Thirty thousand developer accounts in months.

That is agents taking actions with legal and financial consequence, distributed through harnesses, at a scale where the authorisation layer is still an open design question. The monitoring companies are right that this needs watching. Nobody yet knows what they should be watching for.

Axios Pro — Obsidian Security raises at $1.1B valuation → · Google DeepMind — Investing in multi-agent AI safety research → · TechCrunch — Naïve raises $28.5M to automate the grunt work of setting up and running a company →