Yugen AdvisorsYugenAdvisors

AI security for systems with real authority

Secure what your AI systems can actually do.

Yugen Advisors helps product and security teams understand where AI-enabled features, agents, and internal copilots can access data, invoke tools, change workflows, or trigger decisions — then designs the authorization, approval, and audit boundaries around them.

Start a conversationAbout Luis
AI feature reviewsAgent permissionsApproval boundariesAudit-ready evidence
Lighthouse over a dark sea
The mandate: useful AI with clear limits on what it can reach, change, approve, and remember.

Core lanes

Start with the authority boundary.

Most AI risk becomes concrete when you ask one question: what can this system actually cause to happen?

AI feature threat modeling

Map what an AI-enabled feature can read, infer, recommend, approve, invoke, or change before it reaches customers or internal operators.

  • Authority and data-flow mapping
  • Customer-impact and trust-state analysis
  • Risk narratives product and security can share

Agent and tool permission review

Review the tools, APIs, secrets, repositories, tickets, production systems, and business workflows an agent or copilot can touch.

  • Tool scopes and delegated-action boundaries
  • Secrets, identity, and production reach
  • Failure modes, rollback paths, and kill switches

Approval and audit boundary design

Make human review real: define what must be approved, what state reviewers need, what the system records, and where action can stop.

  • Human approval gates that carry context
  • Durable logs and evidence trails
  • Reusable controls for audits and customer trust

What changes after the work

Less fog. More structure.

A clear map of what AI can do

Turn vague AI risk pressure into a concrete map of data access, tool access, approvals, owners, evidence, and next steps.

Bounded AI workflows

Design AI-assisted workflows with explicit permission boundaries, review gates, audit trails, and rollback paths.

Compliance as infrastructure

Convert repeated questionnaires and evidence requests into reusable, reviewed knowledge instead of one-off paperwork.

How we work

Map risk. Build evidence. Make it repeatable.

01

Orient

Understand the product, AI usage, data flows, customer trust pressure, and security maturity.

02

Map

Identify what the AI-assisted system can cause to happen, who owns the risk, and where controls already exist.

03

Build

Create the smallest useful system: permission changes, approval gates, evidence trails, roadmaps, or architecture updates.

04

Operationalize

Make the work repeatable so the team keeps moving after launch, audit requests, or customer diligence.

Writing

Recent notes from the blog.

Short, practical pieces on the security boundary behind AI systems: what they can reach, what they can change, who approves action, and what evidence survives.

2026-07-07

Browser Agents Turn the Web Into an Instruction Surface

When agents browse and act, web pages, documents, ads, and hidden content become potential instructions. Browser automation needs authorization boundaries, not only better refusals.

6 min readbrowser-agentsprompt-injectionai-security
Read post

Why write this way

Decision-useful, not decorative.

Yugen's writing is meant to make the risk legible: what changed, what can now happen, where the boundary failed, and what a team should do next.

Browse all posts

Start here

Bring the messy surface. Leave with a map.

Useful for startups and product teams adding AI features, agents, support automation, or internal copilots before the security model has caught up.

Contact Yugen