Defense in depth from physical hardware through identity-based access and continuous endpoint health monitoring. Each layer addresses different threat classes — no single layer carries the full burden.
23 automated tests aligned to the OWASP API Security Top 10. This validates SQL injection defense across multiple attack vectors, business unit authorization isolation, input validation and sanitization, resource consumption limits, SSRF prevention, CORS policy enforcement, stack trace suppression, and endpoint inventory management.
Every user input passes through regex validation and parameterized queries with bind variables before it reaches a database call. No string concatenation. No dynamic SQL.
17 automated tests aligned to the NIST AI Risk Management Framework across all four functions — Govern, Map, Measure, and Manage. This verifies that every question maps to exactly one approved endpoint, that the AI has no open-ended access to the database, that financial data and general knowledge responses are completely isolated through separate execution paths, and that every AI response is deterministic and reproducible.
53 automated tests across three financial and privacy compliance frameworks. SOX Section 404 (COSO Framework) validates read-only enforcement, segregation of duties, parameterized queries, and result integrity hashing — 17 tests. SOC 2 Type II (Trust Services Criteria) validates access controls, availability monitoring, processing integrity, confidentiality, and privacy — 20 tests. GDPR/Data Privacy validates data residency, data minimization, purpose limitation, transparency, and security of processing — 16 tests.
All queries execute against governed database views — read-only, purpose-built for each reporting function. No INSERT, UPDATE, or DELETE operations are possible through the governed path. Business unit isolation is enforced through an access control table checked on every request. Each business unit sees only its own data.
Every query generates an audit record: the original question, the exact SQL executed, a SHA-256 hash of the result set, a timestamp, and the model version. An independent party can take any AskERP response, recompute the hash from the underlying data, and verify it matches the audit record. Full chain of custody, every query, every time.
Every user is authenticated by email identity flowing from the chat interface through the API layer to the database. Access is controlled at the user + module + business unit level through a governance table (USER_MODULE_ACCESS). A PeopleSoft security sync script maps 30 PeopleSoft roles to AskERP module grants. Users see only the business units and modules they're authorized for. The identity layer includes an in-memory cache, admin endpoints for grant management, and a dry-run mode for the sync process.
Every governed endpoint is automatically tested for availability, data integrity, schema correctness, and response time — nightly. A monthly comprehensive report covers all 70+ endpoints across all modules with row count thresholds, result hashing for drift detection, and cross-endpoint consistency checks. Failures trigger immediate alerts. Monthly reports go to the compliance binder alongside the security test results.
AskERP ships with an automated compliance test suite that generates a self-contained HTML report. Printable, email-ready, designed for CIO and security team review. Tests run nightly at 2:00 AM. The report covers eight compliance frameworks:
23 tests covering SQL injection, authorization isolation, input validation, resource limits, SSRF prevention, CORS, stack trace leakage, and endpoint inventory.
17 tests across all four RMF functions — Govern, Map, Measure, Manage. Validates governed endpoints, three-path isolation, result determinism, and AI traceability.
17 tests validating read-only enforcement, segregation of duties, parameterized queries, result integrity hashing, and internal control monitoring across all five COSO components.
20 tests across all five trust criteria — Security, Availability, Processing Integrity, Confidentiality, and Privacy. Validates access controls, health monitoring, deterministic processing, data isolation, and PII protection.
16 tests aligned to Article 5 principles. Validates data residency (on-premises only), data minimization, purpose limitation, transparency, security of processing, and accountability.
10 tests verifying the end-to-end audit chain — prompt capture, SQL logging, SHA-256 result hashing, timestamp ordering, concurrent safety, and PII exclusion.
Every interaction produces a complete audit chain. Your security team can trace any answer back to the exact query, the exact data, and the exact moment it was generated.
User asks in plain language. The question is captured in the audit log.
Intent classifier maps the question to one governed endpoint. No open-ended interpretation.
The endpoint executes a parameterized query against a read-only database view. The SQL is logged.
Results are formatted for business use. A SHA-256 hash of the result set is computed and stored.
Question, SQL, result hash, timestamp, model version — all recorded. Independently verifiable.
AskERP maintains strict separation between governed financial data, PeopleBooks documentation, and general knowledge. These are three completely different execution paths. They never mix.
User question → Intent classification → Approved service endpoint → database governed view → Structured response with executive summary. Every step is audited. Data comes from live ERP, not the AI model.
PeopleSoft documentation questions — answered through semantic search over locally-stored PeopleBooks PDFs using ChromaDB vector database. Source citations included with every answer. The PDFs, embeddings, and search index all run locally. No external API calls. Clearly labeled as documentation, never mixed with financial data.
Industry benchmarks, best practices, and regulatory context — answered from the local AI model’s training data only. Clearly labeled. Never mixed with financial data. Never sent to a cloud API.
The key point: Your financial data is never used to answer documentation or general questions. PeopleBooks answers come from locally-stored documentation with source citations. General knowledge comes from the AI model’s training data. These three paths never cross.
Security, AI governance, and audit integrity are validated automatically — not assumed. Every release runs through eight compliance standards with a printable report for your security team.
121 tests across OWASP, NIST, SOX, SOC 2, GDPR, and Audit validation run as a single automated suite. One command, one report.
New model versions are evaluated against the full prompt library and compliance suite before they replace the production model. No casual swaps.
Changes move through structured approval paths. The compliance report must pass before any model or prompt change reaches users.
Compliance scorecards support enterprise review and executive visibility. Historical reports show security posture trending over time.
The automated compliance suite covers eight standards in a single run. Each standard maps to a specific threat surface.
Application-layer security against the 2023 OWASP API standard:
AI governance across all four NIST RMF functions:
End-to-end traceability from question to verified answer:
Model changes and prompt updates move through repeatable testing and review — not casual production swaps.
Run the full 121-test compliance suite against the candidate model or prompt change. OWASP, NIST AI RMF, and audit validation in a single automated pass.
Compare candidate output against the governed prompt library — every approved question re-tested for accuracy, format consistency, and governance compliance.
Generate a compliance scorecard: pass/fail/warn per test, remediation notes on every finding, overall compliance status (Compliant / Conditional Pass / Needs Remediation).
The compliance report goes to the security team and stakeholders. No release proceeds without a passing score. Findings are addressed before promotion.
Approved changes are promoted to production with rollback capability. The previous model version is retained for immediate reversion if needed.
Model changes remain controlled, reviewable, and reversible. Enterprise reliability requires discipline at every transition point.
Every model promotion requires a passing compliance report. The 121-test suite must clear before any change reaches production users.
Candidate models are benchmarked against the full prompt library. Accuracy, response format, governance compliance, and performance are measured before any promotion decision.
Every release decision is documented — what was tested, what scored, who approved, what changed. Rollback to the previous model version is immediate.
The compliance suite generates a self-contained HTML report designed for CIO and security team review. No external dependencies. Print it, email it, present it in a meeting.
Monthly scorecards track compliance posture over time. Historical reports show trending — whether the security posture is improving, stable, or has regressions that need attention.
