Frameworks that matter in 2026
- SOC 2 Type II — still the table-stakes ask from US enterprise buyers
- ISO 42001 — the 2024 AI management system standard, increasingly required in RFPs
- NIST AI RMF — US public-sector default, voluntary but influential
- EU AI Act — phased rollout 2025–2027; risk-based classification
- HIPAA — healthcare-specific, requires BAA chain through AI vendors
- GDPR / UK GDPR — Articles 22 (automated decisions), 35 (DPIAs for AI)
Data handling controls
- PII scrubbing on inputs (Presidio, Microsoft PII Detection, custom regex + LLM filter)
- BAA / DPA in place with every AI vendor processing customer data
- Customer data isolation in vector stores (per-tenant namespaces or separate indexes)
- Audit logs of every LLM call: input, output, model, timestamp, user
- Right-to-delete propagation into vector stores and fine-tuning datasets
- Encryption in transit (TLS 1.3) and at rest
- Zero data retention with vendors where possible (OpenAI ZDR, Anthropic ZDR)
Top AI-specific risks
| Risk | Mitigation |
|---|---|
| Prompt injection | Input filters, output validation, scope-limited tools, never trust LLM output as auth signal |
| Data exfiltration via tool calls | Allowlist tools, sandbox execution, output redaction |
| Hallucination in customer-facing output | RAG with citations, confidence thresholds, escalation paths |
| Model abuse / jailbreak | Output filtering, refusal patterns, rate limits |
| PII leakage in logs | Log redaction, structured logging only, short retention |
| Cost runaways (DoS via expensive prompts) | Per-user token budgets, request throttling, cost alerts |
EU AI Act timelines
- February 2025: Prohibited AI practices in force (social scoring, real-time biometric ID, etc.)
- August 2025: GPAI (general-purpose AI) obligations begin
- August 2026: High-risk AI system obligations (most B2B AI sits here)
- August 2027: High-risk AI in regulated products (medical devices, vehicles)
Vendor due diligence checklist
- SOC 2 Type II report (current within 12 months)
- BAA / DPA available and signed
- Data residency options (US / EU / customer-managed)
- Zero data retention option for sensitive workloads
- Training data provenance / opt-out
- Sub-processor list and notification policy
- Incident response SLA
- Right-to-delete and audit-log access
Need an AI security review? Talk to us.