Skip to content
Mirun Kaushik
← All work

07In design2026

TaintTrace

A security layer that stops hijacked AI agents from touching real data

My part
Team lead
Status
In design
Stack
Python, LLM agents, MCP
Code
GitHub ↗

Fig. — How data moves through it

  1. 01Agent tool call
  2. 02Provenance checkwhere did values come from?
  3. 03Policy rules
  4. 04Intent judgeoptional(model call)
  5. 05Real tools or shadow
Model call Rules, data, queues

01Problem

AI agents can be hijacked by instructions hidden in content they read, like an email or a web page. Because the agent acts with the real user’s permissions, normal access checks see nothing wrong.

02Approach

Sit between the agent and its tools. For every tool call, check where the values came from, whether it breaks a rule, and whether it matches what the user actually asked for. Suspicious sessions move to a shadow environment with fake tools and fake data, so nothing real leaks and the agent doesn’t crash.

03What I built

  1. 01I lead the team. I wrote the project reference and the first-phase build plan: the problem, the threat model, the detection checks and the roadmap.

04Not built yet

  • designedThe core engine: provenance and policy checks, shadow tools and session quarantine. Implementation hasn’t started.
  • plannedEvaluation on the AgentDojo benchmark, and an MCP proxy adapter.