AI / LLM Security Testing in Hyderabad · HITEC City

Hyderabad has quietly become one of India's densest AI-engineering clusters — the GCC AI teams inside HITEC City and the Financial District, the ML pipelines behind pharma drug-discovery and clinical operations, and the state's own AI push through T-Hub and the Telangana AI mission. Macksofy runs its South India regional hub out of HITEC City, Madhapur, which means AI/LLM security testing here comes with local senior consultants who can be onsite within two hours across the city — not a fly-in-only model. We test the model, the RAG pipeline, the agent tools and the training-data supply chain, all mapped to the OWASP LLM Top 10 and MITRE ATLAS.

Pharma is Hyderabad's distinctive AI risk surface. Life-sciences companies here are wiring LLMs into clinical documentation, pharmacovigilance triage, regulatory-submission drafting and R&D knowledge retrieval — pipelines that ingest patient data, trial data and proprietary molecule research. Our engagement probes membership inference and model inversion to see whether a fine-tuned model leaks the trial or patient records it was trained on, tests RAG assistants for retrieval of documents the querying user was never cleared to see, and treats PII/PHI leakage as the compliance event it is under the DPDP Act and, for US-facing clinical work, HIPAA.

GCC AI teams in Hyderabad carry a parent company's governance program on their shoulders — Microsoft, Amazon, Google, Qualcomm and dozens of mid-size US/EU centres run real AI engineering here. For those teams we format the engagement so the India centre's evidence maps straight into the parent's NIST AI RMF, ISO/IEC 42001 or EU AI Act high-risk-system controls: a threat model against OWASP LLM Top 10 and ATLAS, per-finding reproducer prompts, and a control-mapping annex the group CISO's team can attest against without a translation layer.

The attack chain is the same rigour we bring everywhere: prompt injection into tool-call abuse into data exfiltration. A crafted document in the RAG corpus or a poisoned support ticket reaches the model as instructions; an agent is steered into invoking a privileged tool on the user's behalf; and sensitive context or system-prompt secrets leak through an output channel the guardrail never inspected. We run the full jailbreak sweep — roleplay, encoding, multi-turn, system-prompt extraction — and test output-format hijacks that smuggle links, markdown and tool-schema injections past naive filters.

RAG and multi-tenancy get direct scrutiny. Cross-tenant retrieval leakage — one user or tenant surfacing another's embedded documents because access control lives in app code the LLM path bypasses — is a failure we test at the vector store's own authorization layer, alongside embedding-space adversarial queries and indirect injection via poisoned corpus documents. For agentic and function-calling systems we fuzz function-call schemas, test confused-deputy tool abuse, enumerate exposed MCP servers, and attempt sandbox escape from code-execution tools.

The supply chain underneath the model is in scope by default: provenance of HuggingFace weights and third-party embedding models, training-data poisoning vectors in fine-tune pipelines, and cost-amplification (long-context and tool-loop) attacks that turn inference spend into a denial-of-wallet problem. For government and government-adjacent AI projects coming out of the Telangana ecosystem, we add the data-sovereignty and audit-trail expectations those engagements carry, and keep the evidence in a form a state audit can accept.

As with every Macksofy AI engagement, we co-build the guardrails rather than just naming the gaps. Each exploitable finding ships with a suggested guardrail prompt, an output-validator rule (PII/PHI, secrets, prompt-leak, toxicity) and a sandboxing or rate-limit pattern the platform team can deploy the same week — with a free guardrail retest inside 30 days built into the base statement of work. Reporting is Mumbai-supported so the HITEC City delivery team pairs local onsite presence with the senior review bench behind every South India engagement.