NeuralTrust's EIC Accelerator Security Gateway for AI Agents

28 October 20261,912 words10 min

In July 2025, two days after xAI released Grok-4, researchers at a Barcelona start-up got it to explain how to make a Molotov cocktail. A month later they did the same to OpenAI's GPT-5 within hours of its launch, as SecurityWeek and CSO Online reported. The company, NeuralTrust, publishes attacks like these to sell the defence: a gateway that sits between a company's AI agents and the models, tools and data they call, and inspects every request.

The EIC Accelerator selected NeuralTrust at its March 2025 cut-off for blended finance. CORDIS records a grant agreement with Singularly, S.L., the company's legal name at the time, for €2,360,423.54, signed on 1 October 2025 and running from 1 November 2025 to 31 October 2027, under the GenAI4EU challenge for European champions in generative AI. The project is one year into a two-year term, so what it promised can only partly be checked.

€2.36MEIC grant recorded in CORDIS
$20MSeed round announced in June 2026
4Customers the company names

Company file

Legal entity
NeuralTrust SL, NIF B10943710, Barcelona commercial register; called Singularly SL until January 2026
Incorporated
July 2022 (deed of 19 July, activity from 20 July)
Registered office
Carrer del Perú 186 bis, Barcelona; offices in London and New York, per the company
Leadership
Joan Vendrell (co-founder, CEO and board chair), Victor Garcia (CTO), Alejandro Domingo (COO)
Investors
Alstin Capital (lead), VentureFriends, Seaya, Kibo Ventures, Banc Sabadell, EA Ventures Plug and Play Fund, Finaves
Products
TrustGate (agent gateway), TrustGuard (runtime protection), TrustLens (agent posture management)
EIC project
Agreement 101247606, 1 Nov 2025 to 31 Oct 2027, blended finance
Staff
Not published

A name change and a new board

The company behind the NeuralTrust brand was incorporated in Barcelona in July 2022 as Singularly SL, with €48,000 of share capital and three joint administrators: Joan Vendrell, Victor Manuel Garcia and Ilan Eskinazi, whose appointment was revoked in April 2024. That is what the commercial register entries published in BORME show, as transcribed by Empresia. The Catalan government's start-up directory also lists Singularly, S.L. as the company operating NeuralTrust, founded in 2022. In December 2023 the company was registered as an empresa emergente, the status Spain's 2022 start-up law gives young innovative firms.

On 20 January 2026 the register entered the change of name from Singularly SL to NeuralTrust SL, together with a move of the registered office from Premià de Dalt to Barcelona. Infoempresa lists NeuralTrust SL as active, with its last register announcement published on 22 July 2026. According to Empresia's transcription, that announcement records the company's conversion from two joint administrators to a board, with Vendrell as chair and chief executive, Garcia, Dertium Capital SL, Georgios Dimopoulos and Alexander Meyer-Scharenberg, the Alstin Capital partner quoted in the seed round release, as directors. It also records two capital increases and one reduction.

The register

Share capital rose by €30,790 in one July 2026 filing

Nominal share capital as entered in the register: €48,000 at incorporation, then two increases and a reduction, all entered on 15 July 2026, a month after the seed round was announced. Nominal capital is not the amount investors paid.

Spanish commercial register entries published in BORME on 27 July 2022 and 22 July 2026, as transcribed by Empresia.

The register shows nominal capital, not the price investors paid for their shares, so it cannot confirm the size of the round. No filed accounts, staff numbers or revenue figures could be found in the free register extracts checked for this article.

How the gateway works

Most defences for large language models work on the prompt: a filter checks what a user types and what the model answers. The CORDIS objective describes NeuralTrust's approach as three linked modules: a Secure AI Gateway that blocks attacks in real time, algorithmic red teaming that probes a customer's AI systems for weaknesses, and observability for monitoring and tracing what the AI does. The company's claim is that it is "the first to secure all LLM traffic at the infrastructure level" and that it analyses the context of user behaviour rather than single prompts. CORDIS also quotes the company as "outperforming market alternatives in public benchmarks", without naming a benchmark.

The products have since been reorganised around AI agents, the software that acts for a user by calling tools and other systems. In its June 2026 funding release, the company describes TrustGate as a gateway that brokers calls to models, to tools through the Model Context Protocol and to other agents; TrustGuard as a runtime engine that detects and stops threats where agents operate; and TrustLens as a tool for finding agents across a company and tracking their behaviour. It says it inspects millions of agent interactions a day and flags about 1.2% as malicious. No independent measurement of that rate, or of how many attacks the gateway misses, could be found.

The attack research

The clearest public evidence of NeuralTrust's technical work is its offensive research. Its Echo Chamber technique, first described on 23 June 2025, steers a model over several conversation turns with harmless-looking prompts until it produces content it would refuse if asked directly. Against Grok-4, Echo Chamber alone was not enough; combined with Crescendo, a technique Microsoft described in April 2024, it succeeded in the share of attempts shown below, according to the figures SecurityWeek took from NeuralTrust's write-up.

The research

Grok-4 gave harmful answers in up to two thirds of attempts

Success rates NeuralTrust reported for its Echo Chamber technique combined with Crescendo against Grok-4 in July 2025, by type of harmful request. The figures are the company's own test results as reported by SecurityWeek, not an independent replication.

SecurityWeek, 12 July 2025, citing NeuralTrust's write-up.

For GPT-5, CSO Online reported that NeuralTrust researchers paired Echo Chamber with a storytelling approach and obtained harmful instructions within hours of the model's launch; the article gives no success rate for that test and reports that OpenAI did not immediately respond. NeuralTrust's own blog post documents the method. The company says Echo Chamber and a later multimodal attack it calls Semantic Chaining are now part of the OWASP AI Security Project taxonomy; that claim was not checked against OWASP for this article.

What the research does and does not show

Breaking a model shows that a team understands how attacks work. It does not show that the same team's gateway stops them. All published numbers on attack success come from NeuralTrust's own tests, and no published third-party evaluation of TrustGate or TrustGuard could be found.

Customers, investors and analyst reports

On 17 June 2026 NeuralTrust announced a $20 million seed round led by Alstin Capital, with VentureFriends, Seaya, Kibo Ventures, Banc Sabadell, EA Ventures Plug and Play Fund and Finaves, the venture fund of IESE Business School. The company calls it the largest cybersecurity seed round raised by an EU company to date, a description Capital Riesgo and Pulse 2.0 repeat. Reports differ on the currency: the company and both outlets give $20 million, Ara converts it to about €17 million, and the IESE alumni newsletter of July 2026 gives €20 million.

The same release names four customers: Air Europa, Abanca, Iberia and Banc Sabadell, which is also an investor. It says 92% of its customers report annual revenue above $1 billion, 80% are in Europe, and that annual recurring revenue in the first quarter of 2026 was double that of all of 2025. No absolute revenue figure is published. The company also announced that KuppingerCole named it a product and innovation leader in its 2025 Leadership Compass for generative AI defence, and that it appears as a representative vendor in two Gartner market guides; at South Summit Madrid in June 2025 it won the awards for most scalable start-up and best trust tech and data start-up.

One detail differs between the company's own pages: the South Summit announcement of June 2025 names Alejandro Domingo as co-founder and CTO, while the June 2026 release lists Victor Garcia as CTO and Domingo as COO.

The record

From Singularly SL to a $20M AI security round in four years

The company was registered in July 2022 as Singularly SL, published its best-known attack research in 2025, signed its EIC grant in October 2025 and took the NeuralTrust name in the register in January 2026.

  1. Jul 2022Singularly SL incorporatedDeed of 19 July 2022 in Barcelona, share capital €48,000, three joint administrators
  2. Dec 2023Start-up statusRegistered as an empresa emergente under Spain's start-up law 28/2022, from 5 December 2023
  3. Jun 2025Echo Chamber publishedMulti-turn jailbreak technique described by NeuralTrust researchers on 23 June 2025
  4. Jul 2025Grok-4 jailbreakEcho Chamber combined with Crescendo two days after Grok-4's release, per SecurityWeek
  5. Oct 2025EIC grant signedGrant agreement 101247606 signed on 1 October 2025; project runs from 1 November 2025€2.36M
  6. Jan 2026Renamed NeuralTrust SLName change and move of the registered office to Barcelona entered in the register on 20 January 2026
  7. Jun 2026Seed roundLed by Alstin Capital with VentureFriends, Seaya, Kibo Ventures, Banc Sabadell and others$20M
  8. Jul 2026Board and capital increasesBoard of directors created and two capital increases entered on 15 July 2026
  9. Oct 2027EIC project endsScheduled end of the grant agreementScheduled

Spanish commercial register entries as shown by Empresia and Infoempresa; CORDIS; SecurityWeek, 12 July 2025; CSO Online, 12 August 2025; NeuralTrust, 17 June 2026.

The EIC money

The award was selected as blended finance, which pairs the grant with a possible equity investment by the EIC Fund. CORDIS records only the grant. The EIC Fund's list of invested companies, checked page by page on 28 October 2026, does not include NeuralTrust or Singularly, and the investors named in the June 2026 round do not include the EIC Fund. Whether an equity investment is still planned is not public. The funding release says the company is also backed by Spain's State Research Agency, without giving an amount; no such award could be found in a web search.

  1. Secure AI Gateway with real-time attack blockingOn saleTrustGate and TrustGuard are sold to named customers, per the company; no independent test of blocking rates.
  2. Algorithmic red teamingShown in publicJailbreaks of Grok-4 and GPT-5 reported by SecurityWeek and CSO Online.
  3. Observability and traceabilityAnnouncedTrustLens described in the June 2026 release; no customer case published.
  4. Outperforming alternatives in public benchmarksNot shownCORDIS names no benchmark, and no third-party benchmark result could be found.
  5. Project resultsToo earlyCORDIS shows no periodic report; the project runs to October 2027.

The European case

NeuralTrust was funded under GenAI4EU, which the European Commission launched in January 2024 to spread generative AI through European industry and public administration. Securing those systems is also a legal requirement: Article 15 of the AI Act obliges providers of high-risk AI systems to make them resilient against attempts by unauthorised third parties to alter their use, including inputs designed to make a model err. A European supplier that banks and airlines can audit fits both aims. The market NeuralTrust sells into is crowded: the same KuppingerCole report, its first on the category, also named F5 a product, innovation and market leader, per F5.

What the company says

The first platform to secure all LLM traffic at the infrastructure level, ahead of alternatives in public benchmarks, used by large European banks and airlines, with recurring revenue doubling and the goal of leading the European market by the end of 2026.

What the record shows

A €2.36 million EU grant running to 2027, a $20 million round, a new board and capital increases in the register, four named customers and widely reported attack research. No independent test of the defence, no published accounts and no EIC Fund investment.

Verdict

A year into its grant, NeuralTrust has raised a large private round and names paying customers. Its reputation rests on breaking other companies' models, and that work is well documented. Whether its gateway stops the attacks it sells against is something only the company has measured so far.

The reporting is based on publicly available sources, listed below.

Industry

Digital and computingCybersecurity

Focus
  • AI security
Technologies
  • Artificial intelligence
  • Cyber threat detection
Keywords
  • Large language models
  • AI agents
  • Prompt injection
  • Generative AI
  • AI governance
  • Financial services