1. OWASP Top 10 v2.0, Agentic Risks (ASI01–ASI10) & Dynamic Benchmarks
The ICO’s October 2026 Agentic AI guidance warns that agent autonomy never excuses data protection breaches: setting permissions makes your organisation a UK GDPR Data Controller. Engineering teams therefore test against OWASP Top 10 v2.0 (elevating Excessive Agency to LLM03, Supply Chain to LLM04, Data Poisoning to LLM05) and OWASP Agentic Security hazards (ASI01 Goal Hijack, ASI02 Tool Misuse, ASI03 Privilege Abuse, ASI06 Memory/RAG Poisoning).
- PINT Benchmark (Prompt Injection Over-Defense): Exposes false-positive over-blocking on benign technical queries; Lakera Guard scores 92.5% with sub-50ms latency (10k req/mo EU free tier), requiring careful FPR threshold tuning.
- HarmBench, AgentDojo & InjecAgent: HarmBench evaluates 510 harmful behaviours by Attack Success Rate, while AgentDojo (629 workspace cases) and InjecAgent (1,054 multi-tool data-stealing cases) prove stateless tests miss indirect RAG/tool injections.
- CI/CD vs Scanner Roles: Use NVIDIA Garak for broad baseline model scanning, Promptfoo (MIT) or Giskard (Apache 2.0, Enterprise VPC ~$2,499/mo) for application regression gates, and Mindgard ($30M Series A, G-Cloud/CSS3) for automated hidden-guardrail reconnaissance.
2. Sovereign London Guardrails, Break-Even Economics & SARIF-to-GRC Wiring
For UK GDPR special-category or zero-retention compliance, AWS Bedrock Guardrails (eu-west-2 London) provides strict zero data retention (30+ PII entity types at $0.10/1k text units; content filters $0.15/1k), while Azure AI Content Safety (uksouth London) offers Prompt Shields against indirect RAG injection at $0.375/1k records (5,000 free/mo). Cost modelling shows cloud APIs beat fixed self-hosted GPUs below ~3,100 queries/day ($131/mo vs $184/mo fixed)—above ~3,100/day, self-hosted VPC models or local regex/token-truncation pre-filters win on margin.
Crucially, raw security logs do not pass ISO/IEC 42001 Clause 9.1 or Annex A.6.2.6 audits until wired into your system of record: export Promptfoo/Giskard/Mindgard results in JSON / SARIF 2.1.0 format straight into Drata Agentic Trust Management REST APIs, Vanta, OneTrust EU AI Act modules, or Holistic AI Sentinel/Operative runtime agents (compare Mindgard vs Lakera).
