Whitepaper · security · 28 min read
Enterprise AI Security & Compliance
A practical security baseline for production agentic AI - covering OWASP LLM Top 10, PII handling, BYOK/HYOK, audit, SOC 2, HIPAA, GDPR and ISO 27001/42001.
A defensible, audit-ready security baseline for enterprise agentic AI. Maps each control to its standard (SOC 2, HIPAA, GDPR, ISO 27001, ISO 42001) and shows how Bhogar AI implements it.
November 26, 2025 · For CISOs - Security Architects - Compliance Leads - GRC
Threat model: what changes with LLMs
Classic AppSec still applies. What is new: prompt injection, training-data poisoning, model supply chain, vector-store leakage, prompt logging containing secrets, agent over-privilege, and emergent behaviours under unfamiliar input. A modern AI threat model addresses all eight as first-class.
Identity and access
Every action runs as a real principal - never as the platform service account. Workspace isolation is enforced at the data, runtime and observability planes. ABAC complements RBAC for capability-scope decisions. SCIM for provisioning, SSO for authentication, OIDC for federation.
Secrets and keys
BYOK as a baseline; HYOK for sovereign tiers. Provider keys live in customer-controlled vaults; rotation is automated and observable; key usage is attributed per agent and per workflow. No raw keys in env vars.
PII and PHI handling
Detect, redact, log carefully. PII is detected at ingest and at every prompt boundary. Redaction policies are tenant-configurable. Vector indexes optionally exclude PII tokens. Logs hash personal identifiers; raw payload retention is minimised and time-bound.
OWASP LLM Top 10 - implementation patterns
For each of the OWASP LLM Top 10 risks (prompt injection, insecure output handling, training data poisoning, model DoS, supply chain, sensitive information disclosure, insecure plugin design, excessive agency, overreliance, model theft) we ship a default control plus an extension point. The defaults are conservative; the extensions are documented.
Guardrails: input, output, action
Three guardrail tiers: input (prompt sanitisation, injection detection, allow/deny lists); output (PII, toxicity, policy compliance, citation enforcement); action (tool-call review, approval gates, blast-radius caps). Bypass is policy-controlled and audited.
Audit logging
Immutable, append-only audit log of every administrative action and every agent decision. Standard fields, queryable, exportable to SIEM, retention configurable per tenant. observability docs has the schema.
SOC 2 Type II
Bhogar AI maintains SOC 2 Type II controls covering security, availability, confidentiality, processing integrity and privacy. Customer-facing controls map cleanly to the trust services criteria; the gap analysis template is shipped with the customer evidence pack.
HIPAA
BAA available for healthcare deployments. PHI handling, audit, encryption-in-transit and at-rest, minimum-necessary access and breach-notification processes are implemented as per the Security and Privacy Rules.
GDPR and DSARs
Lawful basis, purpose limitation, data minimisation, storage limitation and accountability are enforced by design. DSAR fulfilment uses the audit log and the model registry to identify all processing of a data subject’s data; deletion is a workflow with verifiable completion.
ISO 27001 and ISO 42001
ISO 27001 controls map to the control plane (Annex A coverage available on request). ISO 42001 (AI management system) requires policies, model cards, risk assessments and continuous improvement - all first-class concepts in Bhogar AI.
Red-teaming and jailbreak defence
Red-team your prompts. We publish a starter battery of jailbreak prompts plus an automated harness. Defence in depth: multiple guardrails, conservative defaults, refusal logging and metric-driven hardening.
Supply-chain security
Pin model versions, sign artefacts, scan dependencies, monitor for new CVEs, evaluate new providers before promotion. SBOMs available; provenance attestations on request.
Incident response
Documented runbooks for the failure modes in the gateway whitepaper plus AI-specific incidents (prompt-injection campaign, leak via prompt logs, jailbreak chain, runaway cost). Customer comms templates included.
More papers on the whitepapers index.