Designation: AI Safety Researcher (Niche Role) Educational Qualification
• B.
Tech./B.
E. or M.
Tech./M.
S./M.
Sc. in Computer Science, Information Security, AI/ML, or a related quantitative discipline (Must have)
• Advanced degree (M.
Tech./M.
S./Ph.
D.) with a thesis or published work in adversarial ML, AI security, LLM safety, or Responsible AI is highly desirable
• Certifications (Desirable): OSCP, GWAPT or CEH combined with demonstrable AI/ML security work; DeepLearning.
AI or equivalent ML foundations; MLSecOps or LLM-security specialist certifications where available
• Non-traditional backgrounds with demonstrable adversarial ML research, published safety work, credible LLM red-team disclosures, or CTF/red-team achievements will be considered in lieu of formal qualification Experience
• 6+ years total in ML, applied AI, security research, or a closely related discipline; minimum 3 years specifically in adversarial ML, AI red teaming, LLM safety evaluation, or AI/ML security research
• Demonstrable hands-on LLM red-teaming experience with documented prompt-injection, jailbreak, indirect-prompt-injection or agentic-tool-misuse campaigns against production or production-like systems
• Demonstrable adversarial ML work — evasion, model inversion, membership inference, model extraction, or data poisoning — against non-toy classifiers, vision models, or NLP systems
• Prior experience delivering safety, red-team, or Responsible AI work in government, BFSI, healthcare, or another regulated sector is a strong plus
• Prior experience advising or upskilling non-specialist engineering, security, or compliance teams on AI-specific security concerns is desirable

Key Responsibilities
Adversarial ML & Model Security Testing
• Design and execute adversarial attack campaigns — evasion, model inversion, membership inference, model extraction, data poisoning — against pod deliverables including document intelligence, predictive-analytics and fraud-scoring models, and visual AI, face-embedding and liveness models
• Develop and maintain a reusable adversarial ML attack library and test harness for repeatable pre-production evaluation
• Produce adversarial-robustness reports for each pod service with quantitative measures, reproducible attack notebooks, and prioritised mitigation guidance LLM, RAG & Agentic AI Red Teaming
• Lead structured red-team exercises against LLM, RAG and agentic AI deployments across the programme.
• Cover prompt injection, jailbreaks, indirect prompt injection via retrieved documents, data exfiltration, unsafe tool invocation, sandbox escape and policy-boundary violations by agents




• Develop and maintain red-team playbooks tailored to government AI use cases; publish anonymised playbooks and evaluation sets to AIKosh under standard metadata
• Advise pods on guardrail selection, output filtering, RAG source-integrity controls, retrieval provenance, and agent policy design Hallucination, Calibration & Responsible AI Evaluation
• Design and run hallucination measurement, groundedness checks, and calibration/uncertainty evaluation for RAG systems and generative outputs across teams
• Run bias and fairness audits using quantitative metrics — demographic parity, equalised odds, calibration, subgroup accuracy.
• Conduct explainability evaluation (SHAP, LIME, Captum) and lightweight privacy impact assessments on team deliverables
• Author the technical safety evaluation content in model cards, dataset sheets, bias/hallucination/safety evaluation reports, and Responsible AI evidence per the MeitY Responsible AI advisory and the IndiaAI Safe & Trusted AI framework.
• AI Security Architecture & Threat Modelling
• Own AI-specific threat modelling end-to-end for all pod systems under development — STRIDE, MITRE ATT&CK;, MITRE ATLAS and OWASP Top 10 for LLMs — including data pipelines, retrieval sources, model artefacts, prompt paths, tool interfaces and output surfaces
• Review deployment configurations across IndiaAI Compute, on-premise/sovereign GPU and CSP-API paths for AI-specific security risks; recommend controls and defensive patterns
• Contribute AI security and Responsible AI requirements to NeGD RFQs, RDRs and procurement documents; review empanelled-agency proposals for AI security posture Cross-Programme Advisory, Upskilling & Governance Input
• Advise each team’s AI QA Engineer on safety and Responsible AI test design, sample selection, evaluation metrics and evidence capture
• Brief and upskill teams on AI-specific security concerns and Responsible AI controls. Research, Publication & Knowledge Transfer
• Track adversarial ML, LLM safety, agentic-AI safety and Responsible AI research literature; translate relevant findings into team-usable checks, controls and evaluation additions
• Maintain the programme’s open-source AI safety testing toolkit; deposit reusable notebooks, evaluation harnesses and playbooks to OpenForge,



and publish reusable evaluation sets and safety artefacts to AIKosh under standard metadata for reuse Technical Competencies
• Programming & ML Frameworks: Python (advanced); PyTorch or TensorFlow; Hugging Face Transformers; standard data-science tooling (NumPy, pandas, scikit-learn)
• Adversarial ML: Adversarial Robustness Toolbox (ART), Foolbox, CleverHans or equivalent; ability to implement custom attacks and defences; knowledge of certified-robustness techniques
• LLM & Agentic AI Red Teaming: Demonstrable production-relevant experience with prompt injection, jailbreak, indirect-prompt-injection, data-exfiltration, tool-misuse, sandbox-escape and multi-turn manipulation; familiarity with LLM guardrail frameworks (NeMo Guardrails, Guardrails AI, Llama Guard) and open red-team datasets
• Hallucination, Calibration & RAG Evaluation: RAGAS or equivalent; groundedness metrics, faithfulness scoring, retrieval quality metrics, calibration and uncertainty quantification; ability to build custom evaluation harnesses for RAG and generative pipelines
• Threat Modelling & AI Security Frameworks: STRIDE, MITRE ATT&CK;, MITRE ATLAS, OWASP Top 10 for LLMs, OWASP ML Top 10; ability to translate threat models into control specifications and test cases
• Explainability & Fairness: SHAP, LIME, Captum for model explainability; Fairlearn and AI Fairness 360 for fairness metrics; ability to design subgroup-fairness protocols for identity verification, fraud and predictive models
• Privacy-Enhancing Techniques: Working awareness of differential privacy, federated learning, PII redaction and anonymisation techniques; ability to run privacy impact assessments on model and data pipelines
• Government Frameworks & Standards: IndiaAI Safe & Trusted AI framework, MeitY Responsible AI advisory, MeitY Security Policy and Guidelines, CERT-In directions, DPDPA 2023
• Communication & Advisory: Ability to author transparent technical safety reports for a mixed engineering, architecture and executive audience; ability to brief and upskill non-AI-security engineers and compliance colleagues; ability to represent the programme in cross-ministerial working groups Minimum Evidence at Screening (one required)
• Peer-reviewed or workshop-published research on adversarial ML, AI security, LLM safety or Responsible AI
• Documented red-team exercise with verifiable outcomes against a production or production-like AI system
• Open-source contribution to an adversarial ML, AI safety, LLM evaluation, or guardrail project with verifiable commit history
• Prior AI red-teaming, safety, or Responsible AI assessment engagement for a government, BFSI, healthcare or regulated-sector client, verifiable via reference or letter of confirmation

📌 AI Safety Reseracher (New Delhi)
🏢 National e Governance Division
📍 New Delhi