How SigmaAI works — and where its limits are

A security firm\'s AI should be held to the same standard it audits others against. This page explains exactly how our AI processes questions, what grounds its answers, how we control hallucination, and what it will not do.

Where the AI runs

SigmaAI runs on CyberSigma-owned GPU infrastructure. Questions asked on this website are processed on our servers — they are not forwarded to third-party AI providers for routine answering. The one exception is clearly bounded: when our two independent local analyses disagree on a high-stakes internal question, a frontier model may be consulted as a third opinion; public website questions do not use that path.

What grounds the answers

Answers are grounded in a corpus of 2,600+ passages of official texts we are legally permitted to hold in full — the DPDP Act 2023 and Rules, GDPR, DORA, NIS 2, the EU AI Act, RBI directions, SEBI CSCRF, CERT-In directions, the IT Act 2000, NIST publications, CMMC, CCPA, SAMA CSF and the OWASP Top 10 — plus a registry of individually verified facts with sources and verification dates. Copyrighted standards (PCI DSS, ISO, CIS) are represented by verified summaries only: we do not reproduce text we have no licence to hold, and answers about their exact clause numbering say so.

How citations are checked

Every framework citation in an answer is checked by deterministic code — not by another AI — against the source corpus. A clause number the corpus cannot confirm is visibly flagged for verification with an assessor rather than asserted. This is our primary control against the most common failure of AI compliance tools: confidently invented references.

When the AI does not know

Questions outside the corpus are researched against live public sources and the answer is labelled as web-sourced with its links, or the assistant abstains: when evidence is insufficient, SigmaAI is designed to abstain, flag uncertainty or escalate the question for human review.

Human governance

SigmaAI output on this website is informational. Anything that becomes advice to a client — a roadmap, a scope, an assessment position — passes through a senior consultant before it carries CyberSigma's name. Internally, high-stakes AI outputs carry an explicit human-review banner, and AI-cited clause numbers are always confirmed by qualified assessors before client use.

What we store, and consent

Website questions are logged with the question text and an anonymised (hashed) network identifier — no account, no profile. Your details enter our CRM only when you submit them yourself (for example to request a consultant review), and the associated context is stored to make that conversation useful. Standard retention and access rules from our privacy policy apply.

Evaluation and limitations

The stack is tested against an internal benchmark of hard compliance cases — verbatim-accuracy questions, known assessor traps, premise checks and judgment scenarios — and changes ship only when the benchmark passes in full. Honest limitations: model knowledge has a training cut-off and is supplemented (not replaced) by retrieval and live research; long documents are analysed in sections; and no AI output here is legal advice or a substitute for a scoped engagement.

Current evaluation snapshot (15 August 2026 — updated quarterly)

Two evaluation suites run before any model or pipeline change ships. Answer-quality suite: 8 scenario cases across 6 categories — verbatim citation accuracy, assessor-trap detection, false-premise challenge, grounded regulatory detail, response-time discipline, and novel-scenario judgment — currently 8/8 passing. Applicability-calibration suite: 10 cases, all passing — false-mandatory guarding (a third-party payment integration must not yield a mandatory PCI DSS verdict), conditional-status handling (RBI IT Governance Directions apply by Scale-Based Regulation layer, so a generic NBFC gets a conditional verdict with the layer question), regulation-vs-technology-control separation, instrument specificity, SEMANTIC source validation (the cited official document must be the one that supports the specific conclusion — an official-but-wrong document fails the test), phased-commencement accuracy for the DPDP Act and Rules, roadmap chronological consistency, official-source coverage, prompt-injection resistance, and scoping-question presence. Measured properties: 100% of framework citations are machine-verified against the corpus with a flag-or-abstain policy for anything unverifiable (unsupported claims are flagged, never silently asserted); applicability outputs carry per-item confidence and missing-information fields instead of overstated certainty. Corpus version: 2,614 passages, 46 frameworks. Typical response times: simple questions 5-10s; applicability mapping ~15s; deep multi-model analysis 25-130s. All client-facing outputs pass consultant review, and reviewer rejections become new benchmark cases. Suite sizes are published deliberately — a small, hard, fully-disclosed benchmark beats a large opaque one, and both suites grow quarterly.

Responsible AI governance

The same discipline we advise clients on applies to our own stack: documented data-source licensing, tenant isolation on our platforms, audit logs of AI activity, deterministic verification layers around model output, and a standing quarterly review of models against the benchmark before any is adopted.