Why this benchmark exists
Most organizations can describe their cybersecurity posture in detail — vulnerability counts, patch rates, endpoint coverage. Almost none can answer the equivalent question for AI: which AI attack techniques are covered, which threats remain exposed, and how that compares to peers.
The HexTyx AI Security Threat Landscape & Coverage Benchmark closes that gap. Across 21 chapters, it scores enterprise AI security readiness using the HexTyx AI Security Coverage Model™ — six weighted domains spanning prompt security, agent security, RAG security, data protection, supply chain security, and governance — then applies that model across eight industries, six threat categories, and the economics, maturity, and incident response practices behind the numbers.
This Phase 1 edition draws roughly 95% of its data from public frameworks (MITRE ATLAS, MITRE ATT&CK, the OWASP LLM Top 10, NIST AI RMF), governance standards, academic research, and public incident analysis, with the remaining 5% from HexTyx's own assessment data — a mix that will shift toward proprietary telemetry in future editions.
Five things every CISO should know
Agent security is the fastest-growing blind spot
Average coverage of 41/100 — the second-lowest of all six domains — even as autonomous agents gain access to email, CRM, ERP, and financial systems across every industry measured.
AI supply chain security is the weakest domain overall
At 39/100, most organizations cannot fully inventory the models, MCP servers, plugins, and vendors their AI systems actually depend on.
Governance failures amplify every other risk
The majority of major AI incidents analyzed trace back not to a technical flaw, but to a governance failure — no ownership, no inventory, or no approval process — further up the chain.
81% of organizations have never run a formal AI red team assessment
Most AI systems pass conventional penetration tests and compliance audits while remaining fully exposed to prompt injection, agent abuse, and memory poisoning.
Detection speed is the strongest predictor of incident cost
The same AI incident detected within an hour costs an estimated $50K; detected after 30 days, it costs $2M or more — runtime monitoring is now an economic decision, not just a technical one.
The six domains, scored
Every score in this report rolls up to the HexTyx AI Security Coverage Model™ — six weighted domains that make up Tier 1 of the report's three-tier framework architecture.
Coverage by industry
Technology leads every sector measured; retail and ecommerce trail furthest behind, with insurance and legal services occupying a developing middle tier.
| Industry | Coverage Score | Maturity Level |
|---|---|---|
| Technology | 72 | Level 4 |
| Financial Services | 64 | Level 3 |
| Government | 61 | Level 3 |
| Manufacturing | 58 | Level 3 |
| Healthcare | 54 | Level 3 |
| Legal Services | 52 | Level 2 |
| Insurance | 49 | Level 2 |
| Retail & Ecommerce | 46 | Level 2 |
A three-tier framework
The report uses eleven distinct weighted scoring models. Rather than leave that complexity implicit, Chapter 4 introduces a three-tier architecture that organizes every model in the report:
Core Coverage Model
The six-domain, org-wide Coverage Model — Prompt, Agent, RAG, Data Protection, Supply Chain, and Governance — used as the report's primary executive score throughout.
Nine domain-specific sub-rubrics
Deeper weighted models for individual domains — agent risk, prompt injection exposure, data leakage, vendor risk, and more — each scoring one Tier 1 domain in detail.
Executive composite scores
Three board-level metrics — Coverage Score™, Threat Exposure Score™, and AI Maturity Index™ — that synthesize the full report into headline numbers for the boardroom.
All 21 chapters
Explore each topic on the site
Every major theme in this report has a dedicated deep-dive on HexTyx's research hub — read the chapter, then go deeper on the topic that matters most to your organization.
2026 AI Threat Landscape
The full threat taxonomy behind this report's six coverage domains.
Ch. 9 & 18AI Supply Chain Security
The lowest-scoring domain in the benchmark, covered in depth.
Ch. 10AI Agent Security
Permissions, tool abuse, and the agent risk equation.
Ch. 11The State of Prompt Injection
Direct, indirect, retrieval, and multimodal injection types.
Ch. 6 & 12RAG Security
Retrieval poisoning, chunk-level access control, and leakage paths.
Ch. 13AI Governance Failures
The five-layer root-cause model behind most AI incidents.
Ch. 14AI Red Teaming
The seven-phase methodology for testing AI under attack.
Ch. 17AI Incident Response
Detection, containment, and forensics built for AI-native incidents.
Ch. 4 & 20MITRE ATLAS Coverage Benchmark
The interactive tool behind this report's coverage scoring.
Get the full 102-page report
All 21 chapters, the complete Coverage Model methodology, and the full industry benchmark — free to download.
Looking for something else? Browse the full HexTyx Research Hub for the complete library of AI security guides, dashboards, and tools.