Saket Sanjeev Chaturvedi, a postdoctoral appointee, is developing agentic AI frameworks for scientific applications while researching security for large language models, RAG systems, and multimodal perception. His work focuses on making these systems safe and reliable for real-world deployment in research environments.
The dual focus matters for the scientific community. Agentic AI frameworks aim to automate multi-step research tasks, but their usefulness depends on whether researchers can trust the outputs. Chaturvedi's security research addresses what happens when those systems fail or are manipulated.
Two tracks, one goal
Chaturvedi's work splits into two connected areas. The first is building agentic frameworks that can carry out scientific workflows. The second is stress-testing the underlying models - LLMs, RAG pipelines, and multimodal systems - to identify where they break.
Both tracks converge on the same question: can these tools be deployed without introducing errors or security holes into the research process? For labs considering agentic automation, that question is the difference between a useful assistant and a liability.
Why security is a research problem
Security for LLMs is not just about preventing misuse. It involves understanding how prompt injection, data poisoning, or adversarial inputs can corrupt results. In a research setting, a compromised model could produce plausible but wrong conclusions that are difficult to trace.
RAG systems add another layer of exposure because they pull from external sources. Multimodal perception extends the attack surface to images and other non-text inputs. Chaturvedi's research addresses these failure modes before deployment rather than after an incident.
Why this matters for science and research professionals
Researchers evaluating agentic AI tools should ask about the security testing behind them. A framework that automates literature review or experiment design is only as reliable as its resistance to corrupted inputs. Chaturvedi's work suggests that security is not a bolt-on feature but a core part of making AI systems viable for scientific work.
For lab directors and principal investigators, the practical takeaway is to demand evidence of security testing when adopting these tools. The same systems that speed up research can quietly introduce errors if their inputs are not properly vetted.
Your membership also unlocks:
