AI and LLM security testing in India

A classic web test may miss AI-specific paths if the model is treated as a black box text field. AI testing adds adversarial prompting, tool-permission review and data-flow analysis across retrieval and logging. Conversely, AI testing without ordinary authz testing is incomplete — many “AI breaches” are still broken access control on the surrounding application.

• Which models and hosts (vendor API vs self-hosted)
• Which tools the agent may call (email, code exec, tickets, payments-adjacent)
• What corpora are retrieved and how tenancy is enforced
• Whether production data appears in prompts or only synthetic fixtures
• Logging/trace retention and who can read them

Model behaviour is non-deterministic. We report reproducible classes of failure and concrete control gaps; we do not claim exhaustive proof that no prompt will ever succeed. Safety alignment research is not the same as application security testing — we focus on your product’s controls.

Customer-facing copilot on SaaS data. Priority: cross-tenant retrieval and tool actions that modify other customers’ objects.

Internal agent with ticket/email tools. Priority: prompt injection via untrusted tickets leading to data exfiltration or unintended outbound messages.

RAG over mixed public and private documents. Priority: document corpus poisoning and access-control mismatches between search and authorisation.

Enterprise customers increasingly ask how AI features handle their data. Your report should help answer that with specifics: where prompts are stored, whether tenant isolation applies to retrieval, which tools an agent may invoke, and how abuse is monitored. PoCForge findings are written so product and security stakeholders can translate them into roadmap items and questionnaire answers — without claiming certifications we do not provide.

If a buyer’s contract requires CERT-In empanelment for the assessor, treat that as a procurement gate separate from technical scope. Read our CERT-In guide and verify the official list.

Synthetic tenants, sample documents with canary secrets, and tool sandboxes that cannot email real customers make AI testing safer and clearer. Tell us what you can provide in staging. The more realistic the fixtures, the more meaningful the attack paths — without putting production customers at risk.

Treat AI features as another product surface with ordinary access control plus model-specific abuse cases. A sane annual rhythm for many Indian SaaS teams is: core web+API testing on the main product, targeted AI testing when major agent or RAG features ship, and external perimeter checks on a fixed cadence. Use the cost guide to budget the combination rather than shopping for a single magic package.

Large delivery factories can be the right fit for some procurement programmes. Boutique testing suits product companies that want senior attention on their specific architecture, clearer writing, and careful scoping. PoCForge is intentionally small: we publish prices, decline work that is a poor fit, and focus on validated impact. If you need a global brand stamp more than depth, say so — we would rather redirect you than oversell.

Most engagements are remote across India and international time zones. We agree testing windows, communication channels and escalation contacts up front. You receive written findings you can share with engineering and, where appropriate, customer security reviewers. For programme planning across multiple surfaces, start with VAPT services in India and the 2026 cost guide.

If this surface is your priority, send a short brief via contact — assets, environments, roles and timeline. We reply with clarifying questions and a scope-based quote. For broader programmes, combine with other India pages linked above rather than forcing unrelated work into one rushed week.