AI Security · cybersecurity

The cybersecurity of generative AI.

Generative AI adds a new, probabilistic attack surface on top of your existing stack. Here's the threat landscape, the standards that now govern it, and what a credible defence looks like — for enterprises that can't afford to find out the hard way.

A new attack surface

Traditional application security assumes deterministic systems: the same input gives the same output, and a vulnerability, once patched, stays patched. Large language models break that assumption. They take natural language as code, blur the line between instructions and data, and behave probabilistically — so the same attack can fail nine times and succeed on the tenth. Bolt on tools, memory and autonomous agents, and a single successful injection can reach systems and data far beyond the chat window.

That's why AI cybersecurity isn't a subset of your existing programme — it's an extension of it that needs its own threat model, its own testing discipline and its own evidence.

The threat landscape — OWASP LLM Top 10

The most widely recognised map of LLM risk. These are the classes we test for in every engagement:

LLM01Prompt injectionUser input overrides the system's instructions and intended role — directly or hidden in retrieved content.High

LLM02Sensitive information disclosureThe model leaks PII, secrets, system prompts or another user's context.High

LLM05Improper output handlingUnvalidated model output flows into downstream systems — XSS, SQL, command execution.High

LLM06Excessive agencyAn agent with too much tool access takes harmful actions when manipulated.High

LLM04Data & model poisoningTainted training, fine-tuning or retrieval data corrupts behaviour.Medium

LLM08Vector & embedding weaknessesFlaws in RAG pipelines that leak or manipulate retrieved context.Medium

LLM10Unbounded consumptionCost and denial-of-service through unrestricted or runaway usage.Medium

Codes follow the OWASP Top 10 for LLM Applications. Selected entries shown; full coverage is included in an engagement.

The standards that now govern it

AI security is no longer just good practice — it's increasingly a regulatory and contractual obligation. The frameworks that matter:

Regulation · EU

EU AI Act

Risk-tiered obligations for AI systems, with real penalties. High-risk systems need risk management, robustness and human oversight.

Framework · US

NIST AI RMF

Govern, Map, Measure, Manage — the de-facto reference for operationalising trustworthy, secure AI.

Certifiable

ISO/IEC 42001

The first auditable management-system standard for AI — the ISO 27001 of AI governance.

Directive · EU

NIS2 & DORA

Cyber-resilience and operational-resilience duties that increasingly pull AI-dependent processes into scope.

What a credible defence looks like

Adversarial red-teaming — real attacks against the live system, run repeatedly, not a checklist or single try.
Multilingual coverage — guardrails tested in every language you serve, where English-first tooling falls short.
Least-privilege agents — tools, memory and actions scoped so a single injection can't cascade.
Standards-mapped evidence — every finding tied to OWASP LLM, the EU AI Act and NIST so auditors recognise it.
Continuous monitoring — the battery re-run on every model or prompt change, with drift tracked over time.

This page reflects NexusFinLabs' own view of publicly discussed AI-security research and our engagement experience. It is general guidance, not legal advice.

See your AI through an attacker's eyes.

▶ Run a live red-team scan Compliance check 📞