Agentic AI Guardrails

As LLM agents take real actions over infrastructure through protocols such as MCP, safety and reliability become first-class concerns. This work develops policy-driven guardrails that constrain what agents can do, together with automated evaluation frameworks that stress-test agent behavior beyond curated benchmarks.

The goal is autonomous agents that are both capable and provably safe for enterprise IT and infrastructure management.

Representative publications (Kumar et al., 2026; Saranathan et al., 2026).

References

2026

AAAI

InfrastructureSentinel: Policy-Enforced Guardrails for Secure MCP-driven Infrastructure Agents

Tarun Kumar, Aalap Tripathy, Gayathri Saranathan, and 5 more authors

In Proceedings of the AAAI Conference on Artificial Intelligence, 2026

Bib

@inproceedings{kumar2026infrastructuresentinel,
  title = {InfrastructureSentinel: Policy-Enforced Guardrails for Secure MCP-driven Infrastructure Agents},
  author = {Kumar, Tarun and Tripathy, Aalap and Saranathan, Gayathri and Bhattacharya, Suparna and Foltin, Martin and Hinchley, Scott and Bahls, Donald M and Brookshire, David},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume = {40},
  number = {47},
  pages = {40295--40301},
  year = {2026},
}

AAAI-W

Beyond Curated Benchmarking: Automated Evaluation of LLM Agents for Safe and Reliable IT Infrastructure Management

Gayathri Saranathan, Aalap Tripathy, Tarun Kumar, and 9 more authors

In AAAI 2026 Workshop on Agentic AI Benchmarks and Applications for Enterprise Tasks, 2026

Bib

@inproceedings{saranathan2026beyond,
  title = {Beyond Curated Benchmarking: Automated Evaluation of LLM Agents for Safe and Reliable IT Infrastructure Management},
  author = {Saranathan, Gayathri and Tripathy, Aalap and Kumar, Tarun and Hinchley, Scott and Foltin, Martin and Holmes, Christopher L and Brookshire, David and Bahls, Donald M and Xu, Cong and Wisniewski, Robert W and Kaplan, Larry and Bhattacharya, Suparna},
  booktitle = {AAAI 2026 Workshop on Agentic AI Benchmarks and Applications for Enterprise Tasks},
  year = {2026},
}