Skip to content
S&L Labs

Get agents into production with clear boundaries

I review the architecture of agent, RAG and MCP systems and design the changes they need. I build and run systems like these myself.

Worked with

  • Allianz
  • Siemens
  • Heinrich-Böll-Stiftung
  • Universität Bern
  • MAN
  • Hochschule München
  • Lufthansa
  • Hasso-Plattner-Institut
  • Nokia
  • Tactical Tech
  • Evonik
  • Strascheg Center for Entrepreneurship (SCE)
  • Kertos
  • Vitolus

The problem

Prompt injection is unsolved, so the boundaries belong in the architecture

A language model does not reliably separate instructions from data. The German Federal Office for Information Security (BSI) states that no countermeasure removes the problem completely. For a system that calls tools, reads documents and changes data, what matters is what a successful attack can reach.

These are the three places where such systems typically have gaps:

  • Private
    data
  • Untrusted
    content
  • External
    communication

In the intersection: Data exfiltration

Lethal trifecta: If all three meet in one agent, a single crafted document is enough. after Simon Willison (2025)
Permissions
Tools run with the permissions of a service account instead of those of the person asking. The agent then sees more than that person is allowed to see.
Lethal trifecta
Private data, untrusted content and a way to send data out sit in the same context. A crafted document can then be enough to exfiltrate data.
Write access
An agent with write access needs an approval step that shows the change beforehand, and a log that cannot be altered unnoticed.

Services

What I offer

Where I check

  1. Person
  2. Checkpoint: Identity
  3. Agent
  4. Untrusted content Documents, web pages, emails, flows into the agent
  5. Checkpoint: Permissions per tool
  6. Tools and MCP servers
  7. Checkpoint: Approval with diff and log
  8. Data and systems
  • Trust boundary
  • Checkpoint in the review

Fixed price, 5 to 8 days

Go-live review for agents, RAG and MCP

Process, methodology and sample report

Before your system goes live or gets more permissions, I review its architecture and code against the risks in the OWASP Top 10 for Agentic Applications 2026. You get:

  • A threat model for your system, mapped to OWASP Agentic
  • A permission map per tool: which tool reaches which data under which identity
  • A lethal trifecta analysis: where private data, untrusted content and external communication meet in the same context
  • Approval points: where the agent has to wait for a person and what the approval should show them
  • Data flow to the LLM provider: which data leaves your company, where it goes and where it can be masked
  • Targeted attack tests against staging: prompt injection, tool misuse and data exfiltration, manual and automated, with variants of known attacks (fuzzing), each run several times, with reproducible findings
  • For MCP systems: tool poisoning, rug pulls and the supply chain of your MCP servers
  • A report with prioritized findings and a target architecture your team can implement
  • The attack tests as a test suite (evals) that your team runs in CI on every release

1 to 2 days, with your team

Threat modeling workshop

Contents and process of the workshop

Together we model one concrete system, from the point where untrusted content comes in to the last tool call. You get a threat model your team keeps maintaining and a list of the places that need a closer look. This format fits while the system is still being built. If you book a go-live review for the same system afterwards, I credit half of the workshop toward it.

Design and build, scope by agreement

Agent trust architecture

Building blocks and how we work together

I design and build the layer that constrains agents in production, on my own or together with your team:

  • Least privilege for every tool, human in the loop where the impact is high
  • Permissions down to document level, including retrieval
  • Tenant isolation with isolation tests
  • Write approvals with a plain-text diff
  • Audit log from the user’s question to the write operation, stored tamper-evident
  • Tracing of every agent action (OpenTelemetry)
  • Sandboxing for code execution and clean secrets management
  • Masking of personal data before it reaches the model and in test data
  • OAuth 2.1 for MCP servers and external clients

Also without agents

Application security for software your team builds

What is included and what is not

Reviews, workshops and architecture work are also available for applications and APIs without agents. I start where design and code decide security:

  • Security architecture and threat modeling for applications and APIs
  • Authentication and authorization with OAuth and OIDC
  • Tenant isolation in SaaS products, with isolation tests
  • Secrets in code, configuration and deployment
  • Supply chain: dependencies, build and CI/CD
  • Secure code review of the critical paths
  • Vendor dependency and exit readiness: identity, data, keys and models

Standalone penetration tests, network and infrastructure, building an ISMS, migrations and vendor selection are out of scope.

Common stacks

Apps and builders
  • Lovable
  • Bolt
  • v0
  • Cursor
  • Replit
  • Claude Code
  • GitHub Copilot
Agent platforms
  • Copilot Studio
  • Microsoft 365 Copilot
  • Salesforce Agentforce
  • n8n
  • Dify
Frameworks and protocols
  • MCP
  • LangGraph
  • LangChain
  • LlamaIndex
  • OpenAI Agents SDK
  • Claude Agent SDK
  • Vercel AI SDK
  • Haystack
  • CrewAI
Models and platforms
  • Claude
  • OpenAI
  • Azure OpenAI
  • Amazon Bedrock
  • Gemini
  • Mistral
  • Aleph Alpha
  • Llama
  • Ollama
  • vLLM
  • Langdock
Data and operations
  • Qdrant
  • pgvector
  • Pinecone
  • Weaviate
  • Elasticsearch
  • Langfuse
  • LiteLLM
Testing tools
  • promptfoo
  • garak
  • PyRIT

Product names are used for description only and imply no partnership.

Process

How a review works

  1. Intro call

    We clarify which system it is, what prompts the review and how far along it is. You then get a fixed-price proposal with a defined scope and rules of engagement agreed in writing.

  2. Reading

    I read the architecture, code, tool definitions, prompts and configuration. Read access to the repository is enough for this.

  3. Sessions

    I hold two to three sessions with the engineers who built the system. That is where the questions get answered that the code alone cannot answer.

  4. Testing

    I run targeted attacks against the system in a staging environment, with prompt injection, tool misuse and attempts to exfiltrate data. You provide staging access and test accounts. I do not need production data.

  5. Report and target architecture

    You get the findings sorted by impact, each with a concrete proposal for the change. In a closing session I answer your team’s questions.

Practice

From production systems I built

I have built and run the patterns I review, for example at Kertos, a multi-tenant compliance SaaS:

RAG Tenant-separated RAG pipeline
Each tenant has its own knowledge index, synchronized in real time. The agent’s retrieval tool searches only the index of its own tenant.
PII PII masking
Masked data runs through a dedicated export and import pipeline with leak protection and monitoring.
MCP auth OAuth 2.1 for MCP
Implemented: authorization code and device flow, discovery per RFC 8414 and RFC 9728, and approval of external clients per organization.
Approval Declared write operations
Every write by the agent goes through a declared operation with a server-side status check. The approver sees a plain-text diff and approves items one by one or in bulk. Every operation lands in a tamper-evident log with a correlation ID.
Fail-closed Fail-closed approvals
Anything not explicitly allowed stays blocked, even if part of the check fails.
Permissions Object-based permissions
Access rights are object-based and covered by deterministic isolation tests.

These building blocks are in place in several production systems.

What clients say

“Simon designed the architecture around our compliance agent KAIA and brought it to paying customers. He played a key role in establishing evals and made the quality of our AI measurable. Simon takes on far-reaching responsibility, brings the team along and ensures the highest level of security and reliability, even on the most complex tasks.”

Alexander Prams CTO & Co-Founder, Kertos

“As a fractional CTO, Simon conceived and built our first MVP with us, from the data model to the first tests with nursing staff. He brings structure to our ideas and helps us build the right things in the right order. We now develop our AI tools ourselves, with Simon as a sparring partner. When we get stuck, he helps us move forward, and he makes sure security and data protection are in order.”

Moritz Dillmann Co-Founder & CEO, Workinger

Both quotes were given in German and are translated here.

Scope

Scope and limits

I review how your system is built on the inside, test it at the points where agents are vulnerable and design the changes. A penetration test covers application and infrastructure broadly, and a compliance audit checks whether evidence is in place. Guardrails filter inputs and outputs. The boundary still belongs in the architecture.

Applies to every engagement
  • Targeted tests. I attack your system where agents are vulnerable: against staging and within rules of engagement agreed in writing. This does not replace a penetration test of your infrastructure. If you need one, I describe its scope for you.
  • No legal advice. Whether NIS2 or the AI Act applies to you is for your lawyers to decide. What follows from it technically is part of my work.
  • No certification. The report is an expert assessment. It does not come with a seal of approval.
Simon Liedtke

About me

Simon Liedtke

I am a software architect based in Munich and have been building software for almost 20 years. Since 2025 I have mostly built agent systems, MCP servers and RAG pipelines for production use, including evals, LLM observability and the permission and approval layer this page is about. I work independently, so you talk to the person who reviews your system.

  • Fixed price
  • I do the work myself
  • At most three engagements at a time

Technical writing

Writeups

Contact

Talk about your system

Book an intro call or write to me about your system and your timeline. I reply to emails within two business days.