AI is speeding ahead so quickly that the business world hasn’t entirely caught its breath, and you can almost feel the tension between opportunity and exposure. The numbers paint the picture in stark strokes: more than three-quarters of organizations now run AI somewhere in their operations, up from barely half two years ago. That kind of acceleration usually leaves a trail of unguarded doors, and sure enough, the Thales Data Threat Report suggests that nearly three-quarters of those same organizations are now pumping money—sometimes newfound, sometimes carved from old budgets—into AI-specific security. A quiet admission that AI’s promise has arrived hand-in-hand with new and unfamiliar risks.
Against that backdrop, Thales is stepping forward with what amounts to a structural attempt at redefining the security perimeter: the AI Security Fabric. The idea feels a bit like pulling together loose threads of a fast-expanding ecosystem and stitching them into a controlled environment where data, identity, and model behavior can be monitored without suffocating innovation. The goal is deceptively simple—give enterprises enough confidence to scale LLM-powered applications without stumbling into the usual traps of prompt manipulation, data leakage, or unaccounted-for model behavior.
What Thales is rolling out now are the first load-bearing beams of this fabric. The AI Application Security layer acts almost like a specialized WAF for LLM-native apps, watching the traffic that most organizations still treat with unease. It detects injection attempts, jailbreaks, system prompt exposure, malformed input designed to exhaust compute—that whole messy universe of attacks that didn’t exist five years ago and now hit production systems weekly. Deployment looks intentionally flexible, acknowledging that enterprises are no longer neatly categorized: some run cloud-native stacks, some cling to on-prem boxes, most live in hybrid purgatory.
RAG Security is the other early cornerstone, tackling the spot where most enterprises quietly accept risk: whatever lives inside their knowledge bases. Before that data reaches any AI pipeline, the Thales stack wants to fingerprint, classify, encrypt, and manage keys with the kind of discipline normally reserved for regulated sectors. The same system then secures the back-and-forth between models and external data stores, so the retrieval layer doesn’t unspool into an unintended leakage vector. It’s one of those areas where an organization only discovers the danger after the fact—usually after an LLM rephrases something that really should have stayed locked down.
Sebastien Cano’s framing—security designed specifically for Agentic AI and GenAI workflows—captures the broader shift. IT leaders aren’t just containing threats; they’re learning to defend dynamic systems that generate, ingest, and reinterpret data in ways traditional architectures weren’t built to understand. A static firewall or fixed DLP workflow simply doesn’t map to systems that evolve with each interaction.
Looking ahead to 2026, Thales is already sketching the next pieces of the puzzle. Data leakage prevention tuned for model interactions, a dedicated MCP security gateway to audit and control every agent-model-data exchange, and a unified layer of runtime access control across those flows. It feels a bit like the beginning of a new category: not AI-security-as-addon, but AI-security-as-infrastructure, the connective tissue that will sit beneath future enterprise systems the way identity management or network segmentation did in earlier eras.
You can sense the direction: as AI becomes a core operating layer of the enterprise rather than a bolt-on capability, the systems guarding it need to evolve with the same level of contextual awareness. Thales is positioning itself as one of the early architects of that transformation, offering a framework that tries to keep pace with a technology that refuses to slow down.