DeepKeep's EIC Accelerator Red Teams AI Before Attackers Do

The security problem with generative AI is that the same input can be safe today and a leak tomorrow. DeepKeep, founded in Tel Aviv in 2021 by Rony Ohayon, sells a platform for that: a runtime firewall with guardrails, adaptive red teaming, visibility and policy control over how staff and developers use AI, a scanner that maps every tool and action an agent can reach, and static and dynamic scanning of models in the supply chain.

The EIC Accelerator selected the company at its October 2024 cut-off for blended finance, and CORDIS records a grant of €2,490,250 from June 2025 to May 2027 to extend the platform beyond language models to computer vision, spatial sensing, human-machine interaction and multimodal models, including threats that arise when AI is wired into physical environments, with the platform able to run on edge devices and to assess regulatory compliance. The grant agreement was signed on 3 October 2025.

A fourth company and a market that appeared

Rony Ohayon holds a doctorate in communication systems engineering and founded LiveU and DriveU before DeepKeep. The company came out of stealth in May 2024 with a $10 million seed round led by Awz Ventures, the Canadian-Israeli fund, as SecurityInfoWatch reported, saying it was already deployed by global enterprises in finance, security and AI computing. It has named none of them.

Its visibility since has come from analysts and awards rather than customer announcements: Omdia listed it as a key innovator in AI security in November 2025, Gartner named it a sample vendor in its emerging tech impact radar for the AI cybersecurity ecosystem that month and included it in a 2026 competitive landscape on AI operations management, and it won a Globee gold award and a Hacker News award in 2026. The company publishes research too, including work on adversarial detectors on arXiv and an analysis of security flaws in Llama 2 that was picked up in 2024.

What the case studies show

DeepKeep publishes two case studies from its Vibe AI Red Teaming product, both anonymised. In the first, a European cloud provider's support agent was tested unguarded and then with open-source guardrails. Unguarded, 13 attacks produced 15 reported policy breaches; with guardrails, 20 attacks produced 11. The findings in both configurations included data disclosure, unauthorised account actions, persistent prompt injection and abuse of internal tools.

In the second, a food-service chain's voice-activated ordering kiosk, built for blind customers, was tested before production: 16 scenarios across four topics produced six failures, the most serious being that declared allergies and dietary restrictions were recorded for recommendations but not enforced when items went into the cart.

In July 2026 the company also published a benchmark claiming F1 scores approaching 0.98 for prompt injection detection across 12 additional languages, and better results than Meta's Prompt Guard and Nvidia's NeMo on prompt injection and personal data detection, using a classifier of about 400 million parameters that reads meaning directly instead of translating to English. The study is the company's own, and no per-language breakdown or independent replication has been published.

1. Commercial platformOn saleListed on AWS and Google Cloud marketplaces, February 2026.
2. Agent and multimodal coverageShippingAgent scanner and Vibe red teaming launched in 2026.
3. Independent validation of detection claimsNot publishedBenchmark is the company's own.
4. Disclosed customers and revenueNoneCase studies and testimonials are anonymised.

A seed round and a grant

The $10 million seed of May 2024 and the €2.49 million EIC grant are the only funding DeepKeep has itself announced. The project's business model is annual licences in three tiers, cloud, on-premise and edge, with optional purchase of LLM tokens, and it targets 30% of the trust, risk and security management market for AI. No revenue, customer count or new financing round is public, and the company's accounts are not published.

European regulation as the sales argument

DeepKeep's own marketing leads with European rules: the EU AI Act, GDPR and NIS2 all demand evidence an enterprise can produce, and the company publishes guidance on what NIS2 articles 20, 21 and 23 require for AI governance. Its research has also focused on a specifically European problem, that guardrails which translate to English before judging a prompt lose accuracy in German, Spanish, French and Italian. That fits the EIC's interest in AI tools built for the European market, and the grant's own objective includes assessing regulatory compliance.