Methodology
Compiled from 40+ SOC 2 Type II, HIPAA, and GDPR audit engagements delivered between 2023 and 2026 for companies running production LLM-based systems. We cross-referenced auditor questionnaires, observed remediation patterns, and documented the controls that survived the final report. Contributions came from security engineering leads at companies ranging from Series A SaaS to regulated healthcare platforms.
Key findings
The signals that shaped the rest of the analysis — each one backed by the full methodology in the download.
78%
Of AI systems fail their first prompt-injection red team
Including systems that had already passed a formal security review.
12
Controls auditors ask about almost every time
We documented each one with acceptance criteria and evidence templates.
3.1x
More findings for systems without model cards
Formal model documentation reduced audit findings by nearly a third across the sample.
56 days
Median SOC 2 Type II gap closure for AI systems
Versus 38 days for conventional SaaS — the delta is almost entirely data-governance work.
92%
Lacked per-inference audit trails at kickoff
The single most common remediation demand across regulated customers.
Inside the report
A scannable chapter guide. Each chapter includes data tables, charts, and practitioner commentary in the full PDF.
Prompt injection, data exfiltration, model poisoning, and supply-chain risks mapped to concrete controls.
Retention, redaction, and purpose limitation requirements under GDPR and HIPAA, with template DPIA.
The 12 auditor-standard controls, with acceptance criteria, evidence, and example policy language.
BAA coverage, PHI handling, and de-identification pipelines that survive audit scrutiny.
Article 22 obligations, human-in-the-loop design, and right-to-explanation implementations.
Per-inference logging, prompt/response retention policies, and PII-safe telemetry patterns.
How to assess OpenAI, Anthropic, AWS Bedrock, and Azure OpenAI under a formal vendor-risk program.
From first prompt-injection report to root-cause remediation and disclosure timelines.
Full report
All the data, charts, and appendices — no gate, no drip campaign, no retargeting pixel circus.