Skip to content
View Jiten76's full-sized avatar

    Block or report Jiten76

    Block user

    Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

    You must be logged in to block users.

    Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
    Report abuse

    Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

    Report abuse
    Jiten76/README.md

    Jiten Joshi — security engineering for the AI stack

    I build the controls an LLM or agentic system needs before you point it at production: posture, guardrails, task-scoping, and runtime governance — scoped to only the AI surface (Bedrock-gateway AWS, Vertex, Azure AI, on-prem models), not a generic tool retrofitted onto it.

    The bet: AI systems need what cloud workloads got from Wiz and dependencies got from Endor — a security layer built for their actual attack surface. That surface is now well-defined, and I build directly against it:

    What I've built

    agentic-risk-portfolio · live demo → Three working tools on one platform: what a bad year costs (priced risk, correlated Monte-Carlo), an AI agent held to a tamper-evident evidence chain, and minute-by-minute incident replay. Every number is seeded, sealed, and re-checkable in the browser. Findings crosswalk to a canonical KRI catalog and export to OSCAL.

    Yukti Agent-fleet governance. Catches an agent doing what it was never sanctioned to do — unsanctioned egress, excessive agency, identity fabrication — judged from the trace against a per-run task envelope, so it holds even when the model's own guardrails are off. Grounded in the real 2026 agent-eval incidents (UK AISI, OpenAI, Anthropic). Five governance domains on one engine; OTLP trace ingest; consequence-ranked, human-in-the-loop decisions on a self-verifying evidence ledger. 156 tests, CI, mypy --strict.

    The through-line

    Governance you can act on, not just observe. Read from the traces AI systems already emit → detect behavior that crosses a sanctioned boundary → rank by consequence → escalate to a named human on an SLA clock → write it to an evidence chain that proves itself. Observability tells you what happened; this decides what to do about it.

    I'm building toward one AI-native security platform that unifies these into posture + supply-chain + data-flow + agentic-runtime governance for the AI stack — the way a CNAPP does for cloud, scoped to AI alone.

    MBA · focused on AI/LLM security · open to conversations.

    Popular repositories Loading

    1. agentic-risk-portfolio agentic-risk-portfolioPublic

      What does a bad year cost — and can you prove the number? Which shared dependency drags everything down at once? Did an AI agent just touch a tool it never should have? How fast do you really detec…

      HTML

    2. Jiten76 Jiten76Public

      AI-stack security engineering

    3. yukti yuktiPublic

      Agent-fleet governance: read traces, detect behavior crossing a sanctioned boundary, rank by consequence, and act — across five domains, with a tamper-evident evidence chain.

      Python