We are handing real decisions to systems that can't prove what they did.
Shaik Ahamed Sahir. I work on the security of autonomous AI agents: tamper-evident audit trails, how multi-agent systems collude when told not to, and attacks that silence the agents built to catch them.Masters in CS @UMass Amherst & Bachelors in CSE from Sri Venkateswara University
About
I came to security research from building. The question that moved me was narrow and wouldn't close: when an autonomous agent tells you what it did, what makes that account trustworthy? Not persuasive; verifiable. Most agent logs today are written by the same system they're meant to hold accountable, which makes them a record of intent, not evidence. That gap is what I work on: cryptographic audit trails that survive an adversary, and a clear-eyed account of how agents behave when they think no one can check.
Stack
Work
the tools, the papers, and the work that led to both.
Research
AgentOps Replay
Event-sourced backend capturing every agent execution step, with real-time replay, any failure reproducible in under 2 clicks.
FastAPI · React · PostgreSQL · AWSView DOI ↗Tesserae & FACADE
A cryptographic audit layer for multi-agent systems (Tesserae), paired with a taxonomy of exactly how agents collude even after you tell them not to (FACADE).
Rust · Ed25519 · Hash chainsVIABLE
An attack that makes AI security agents silently stop searching the web: without tripping any of the five detectors built to catch it.
Adversarial ML · AI Security AgentsBuild
Your AI agent logs are not evidence. Sasana makes them defensible.
Rust · SHA-256 · Hash chainsGitHub ↗Shyniq
An e-commerce storefront for jewelry cleaners: real products, real orders, shipping today.
Next.js · React · TailwindCSSVisit site ↗Loom
Planner, coder, and reviewer agents that solve subtasks in a sandboxed executor and share long-term memory across sessions, verified by a hash-chained ledger, not self-reported pass/fail.
Python · FastAPI · LangGraph · Qdrant · DockerGitHub ↗Open Source
Fixed bullet/ordered list rendering in rich text shapes
Root-caused to tiptap's ProseMirror layer after review cycles with the founder and core maintainers.
2 merged fixes in typography and markdown serialization
Pipe characters inside inline code spans no longer break markdown table cells; plus opt-in RTL smart quote pairing for Arabic, Hebrew, and Urdu, with zero change to existing LTR behavior.
7 merged fixes across selection, tables, Yjs sync, DnD compliance, and the playground
Line breaks inside slots that couldn't be inserted or deleted, table alignment breaking in non-standard selection directions, Yjs desync after clearing nodes, selection growing mid-collaboration, DecoratorTextNode formatting, dragover spec compliance, and a playground colour modal that dismissed on first click.