HiddenLayer’s $100M Series B: AI Security Goes Mainstream
TechCrunch reported on 2 September that HiddenLayer closed a $100 million Series B led by Delta-v Capital, with Ten Eleven Ventures, Morgan Stanley, Microsoft’s M12 and Booz Allen Hamilton participating. The headline number is not the round — it is that annual recurring revenue grew more than 10× in a year, over 90% of it from new customers. Enterprises are not piloting agents any more; they are deploying them and buying protection.
- Published
Primary source: TechCrunch, 2 Sept 2026. Figures below are as reported there; our analysis follows.
What was announced
| Item | Detail |
|---|---|
| Raised | $100M Series B |
| Lead | Delta-v Capital |
| Also in | Ten Eleven Ventures, Morgan Stanley, Microsoft M12, Booz Allen Hamilton |
| Product | Discovery, runtime protection, attack simulation and supply-chain security for models, agents and workflows; scans ~50 AI file frameworks |
| Threats covered | Prompt injection, agent manipulation, malicious tool use, supply chain |
| Traction | ARR up 10×+ in a year, "tens of millions"; >90% of growth from new customers |
| Customers | Financial services, large tech, U.S. DoD and intelligence community; a frontier model provider with 700M+ weekly users |
| Use of funds | Sales and distribution, engineering and research, expansion into Europe/EMEA |
How to read it
- Runtime is the new perimeter. HiddenLayer’s CEO framed the product to TechCrunch as endpoint detection and response for AI — protection while the agent runs, not just a scan before deploy.
- 10× ARR from new logos means the buyers are new to the category. Security teams that had no AI line item last year have one now.
- The threat list — prompt injection, agent manipulation, malicious tool use — is the same list a good agent harness defends against by design. Buying a product and building it right are complements, not alternatives.
The runtime controls every agent should have
| Control | What it stops | Cost to add |
|---|---|---|
| Input treated as untrusted (documents, tool results, web) | Prompt injection via content | Design decision, ~free |
| Approval gates on irreversible actions | Manipulated agent doing damage | 1–2 days |
| Allow-listed tools with scoped credentials | Malicious tool use, privilege creep | 2–3 days |
| Full call tracing + anomaly alerts | Silent compromise | 2–4 days |
| Adversarial eval set in CI | Regressions that reopen a hole | 1 week |
That table is roughly two weeks of work on an existing agent, and it is what we install first when a client asks us to harden something already in production. If a vendor product comes later, it lands on a system that was built to be defended.
Frequently asked questions
Is AI security only an enterprise problem?
The attackers do not check company size. Any agent that reads external content and can take actions needs the basics in the table above.
Do these controls slow the agent down?
Tracing and allow-lists add negligible latency. Approval gates add human time by design — scope them to irreversible actions only.
Can you audit an agent we already run?
Yes. A one-week review produces the tool inventory, the gaps against the table above, and a fixed-price plan to close them.
