NeuralTrust Secures $20 Million for Enterprise AI Agent Security

Most enterprises deploying AI agents at scale have almost nothing securing them. NeuralTrusta startup operating out of Barcelona and New York, has raised $20 million in seed funding to change that. Alstin Capital led the round, with VentureFriends, Seaya, Kibo Ventures, Banc Sabadell, EA Ventures Plug and Play Fund and Finaves joining. NeuralTrust claims it is the largest cybersecurity seed round raised by an EU company to date, though that claim is self-reported at this stage.

Why AI Agent Security Is a Distinct Problem

AI agents are not just another class of software to secure. They make decisions, execute workflows and interact with sensitive data at machine speed, often across multiple tools and systems at once. That autonomy dramatically expands the attack surface. A compromised or misconfigured agent is not a broken application sitting idle, it is a moving threat actor with API access.

Prompt injection gets the most attention: malicious instructions embedded in agent inputs that override intended behaviour, but the threat profile runs deeper. Agents can leak data through overly broad permissions. Training data can be manipulated to produce flawed outputs. Compromised API keys or OAuth tokens enable identity-based attacks. In multi-agent orchestrations, failure modes compound quickly, over-privileged agents inherit excessive access, data flows unintentionally between agents, and a single misconfiguration can cascade through an entire workflow.

There is also the shadow AI problem. Employees using unsanctioned AI tools create an execution layer that sits entirely outside formal governance, with sensitive decisions and data transformations happening with no visibility and no controls.

Policy, Runtime and Posture: NeuralTrust’s Technical Approach

NeuralTrust’s architecture spans three products, each targeting a distinct layer of the agent security problem.

TrustGate is the gateway. It sits at the intersection of every large language model (LLM) call, Model Context Protocol (MCP) interaction and tool invocation, intercepting, inspecting and enforcing policies before actions reach target systems. If an agent attempts something outside its defined permissions, TrustGate is the enforcement point.

TrustGuard handles runtime. AI agents generate and execute code dynamically, and their behaviour shifts based on context, conversation history and retrieved data. Traditional security tools are built for deterministic software and do not handle that variability well. TrustGuard is designed to detect anomalies and block threats in real time: remote code execution attempts, credential theft, data exfiltration. Detection happens inside the live environment, not after the fact.

TrustLens is the posture layer. It discovers and inventories agents across the enterprise, then continuously monitors the gap between declared permissions, what each agent is supposed to be able to do, and observed behaviour, what it actually does. That gap is where exploitable risk lives. TrustLens also tracks AI supply chain exposure and maintains an up-to-date inventory of agent deployments, effectively an AI bill of materials.

Bridging the Governance Gap

Traditional IT governance assumes predictable, rule-based systems with fixed permissions. AI agents break that assumption immediately. They are adaptive, tool-enabled and capable of unscripted decisions. The governance frameworks built for conventional software do not transfer.

NeuralTrust’s argument is that policy enforcement needs to happen at the infrastructure layer, not in model-level guardrails or written policy documents that go stale within months. By combining policy, runtime and posture management in a single platform, the company lets organisations define granular rules, which data agents can access, how information is combined, which actions are permitted, what outputs can be produced, and enforce those rules in real time rather than auditing them after the fact.

The practical payoff is auditability. When an AI agent makes a consequential decision, TrustGate and TrustGuard together produce a traceable record linking that decision to a defined policy and a human role. That matters for compliance, and it is increasingly a legal requirement. For teams working through similar accountability questions, the Congressional push to codify human responsibility for AI decisions is worth tracking.

Where Current Solutions Fall Short

The AI security market is fragmented, and existing tools leave significant gaps. Traditional application security products are built for static code analysis in CI/CD pipelines, they have no mechanism for monitoring dynamically generated agent behaviour at runtime. Cloud security posture management (CSPM) and data security posture management (DSPM) tools secure the underlying infrastructure and data stores, but cannot tell you which agent accessed a vector embedding derived from sensitive health records, or why. A CSPM detects the open S3 bucket; it does not understand what was in it or who put it there.

AI Security Posture Management (AI-SPM) is emerging as a category, but many early products are essentially relabelled CSPM dashboards: inventory and configuration without runtime behavioural analysis. Broader security platforms from Palo Alto NetworksMicrosoft and SentinelOne address AI-related threats, but within larger cloud and network security frameworks. Specialist players like Zenity, Reco AI, Lasso Security and GuardionAI cover narrower ground: intent-aware governance, shadow AI discovery, action monitoring. NeuralTrust’s claim is that none of these deliver all three layers, policy, runtime and posture, in a single platform.

Regulatory Pressure and the Drive for AI Accountability

Regulation is strengthening the business case. The EU AI Act sets explicit requirements around data provenance, model accountability, explainability and human oversight, and it applies directly to the kind of autonomous agent deployments NeuralTrust is targeting. For enterprises operating in or selling into the EU, compliance is not optional. The August 2026 transparency deadlines under the Act are approaching fast.

The National Institute of Standards and Technology (NIST) AI Risk Management Framework also provides structured guidelines for managing agentic AI risks. For CISOs, the compliance picture is genuinely complex. Explainability requirements, audit trail mandates and evolving bias and fairness standards all require technical controls that can produce evidence, not just policies written in a document. NeuralTrust’s audit trail capabilities are designed to produce exactly that kind of verifiable record.

Building Trust into the AI Agent Lifecycle

Enterprises are already deploying AI agents in finance, procurement, customer support and development workflows. The question is not whether to trust them, deployment decisions have already been made. The question is whether that trust is earned through verifiable controls or simply assumed.

TrustLens establishes the baseline: a continuous inventory of what agents exist, who owns them, what data they access and what their expected behaviour looks like. TrustGuard monitors for deviations from that baseline in real time. TrustGate makes policies executable rather than aspirational. Together, the three products are designed to produce a verifiable record of agent behaviour, the evidence layer that makes auditing and accountability possible rather than theoretical.

Whether the platform delivers at enterprise scale remains to be seen. NeuralTrust is at seed stage, the product claims are largely self-reported, and the shift from reactive post-incident review to continuous behavioural monitoring is architecturally ambitious. The market need is real; execution is the open question.

What To Watch: The Maturation of Agent Security

NeuralTrust’s round is an early data point in what will be a larger market. The most immediate shift will be AI-SPM tools moving beyond inventory into real-time behavioural analytics. Static posture assessments are already insufficient for agents that change behaviour mid-session; the market will consolidate around platforms that can adjust policies dynamically based on observed risk.

Standardisation is lagging behind deployment. NIST, the Cloud Security Alliance and OWASP have each published frameworks, but they do not yet add up to a unified, actionable standard for autonomous agent runtime controls. Large enterprise adoptions and inclusion in vendor certification programmes such as ISO 27001 will be the signal that the field is stabilising.

Identity and access management (IAM) also needs to evolve. AI agents are non-human entities that authenticate, access systems and execute decisions, often with permissions scoped for a human worker. Managing and authorising those non-human identities with the same rigour applied to people is an unsolved problem at scale.

The shadow AI challenge will not resolve itself. Discovery tooling that can surface unsanctioned agents across an enterprise environment and route them into centralised governance will become a standard procurement requirement as agent counts grow.

Finally, watch the cyber insurance market. Autonomous agent deployments introduce a risk profile that existing underwriting models were not built for. New pricing structures that specifically account for agentic AI will create a direct financial incentive for verifiable security controls, and may end up being a stronger adoption driver than regulatory pressure alone. For more coverage of AI chips and infrastructure, visit our AI Hardware section.