Services
Find out what your AI agent can be tricked into doing, before your users do
I review and harden AI agents and MCP servers, and build small agents with the guardrails already in. Start with a free preflight; every paid step below links to the public repositories and articles it is drawn from, not to a claim.
Starting prices in USD. Final scope and price are agreed in writing after a short call, before any work starts.
Pilot pricing: the first three paid clients get 30% off in exchange for permission to publish an anonymized case study.
-
00 / Services
Free MCP / Agent Security Preflight
Free · Usually within a week
Who this is for
Anyone with a public MCP server or AI agent repository who wants a second pair of eyes before launch, and wants to see what a review finds before paying for one.
The problem
Most teams cannot tell whether a security review is worth paying for until they see what one actually finds on their own code.
What you get
- A manual trace of every tool parameter that reaches a file write, a URL fetch or a shell command, checking for a boundary on each.
- An mcpscan run as a second pass: 30 rules mapped to the OWASP MCP Top 10. Its output is triaged by hand, because a scanner alone misses bugs and raises false alarms.
- The top three risks, each with file, line and why it matters.
- A short email: what I found, what I did not check, and whether a full review makes sense for you. Sometimes the honest answer is no.
Not included
- Private repositories, live systems or production endpoints. Public code, or code you choose to share with me, only.
- Behavioral testing, fixes or a formal report. Those are the paid review.
Proof of work
Send a public repository link and get the top three risks I see, free.
Start this conversation → -
01 / Services
AI Agent / MCP Security Review
From USD 1,500 · About 1 week
Who this is for
Teams shipping an MCP server, an AI agent with tool access, or a Claude/Copilot-integrated product, especially before exposing it to other users or tenants.
The problem
A happy-path demo is not evidence that an MCP surface is safe. Write-side path traversal was behind four real 2026 MCP CVEs, including a CVSS 9.1 unauthenticated RCE chain. Tool poisoning, tenant isolation gaps and CI token reuse are the other bug classes a happy-path demo never exercises.
What you get
- A manual source review of every tool handler, tracing caller-controlled input to file writes, outbound requests and shell calls.
- A static scan of your MCP manifests, tool definitions and .claude/ directory with mcpscan (30 rules), triaged by hand: tool poisoning, command injection, path traversal, SSRF, secrets and CI/CD workflow risks.
- Behavioral tests for tenant isolation, write-tool approval and authorization, run against a fixture you control.
- A report with rule ID, severity, redacted evidence and remediation, so each finding is an engineering task, not a vague warning.
- One scoped re-test after fixes ship.
Not included
- Testing production systems without written authorization from the owner.
- Fixing the code. Every finding comes with remediation guidance; the fix is your team's, or a separate sprint.
- A guarantee that nothing else is wrong. A review lowers risk, it does not certify that a system is safe.
Proof of work
Send the repo link for your MCP server or agent tool, and I will tell you what a scoped review would cover before we talk about price.
Start this conversation → -
02 / Services
Agent & Security Automation Build Sprint
From USD 2,500 · About 2 weeks
Who this is for
Small security, IT or operations teams with one repetitive process, such as alert triage, access reviews or ticket intake, that an agent could take over while a human still approves the steps that matter.
The problem
Most agent pilots stall because nobody scoped one process, set the guardrails, or measured the before and after. Manual triage does not scale past a handful of tickets a day, and free tools catch less than teams assume: a default Semgrep scan reported 3 of 10 planted vulnerabilities in one test.
What you get
- One agent or automation for one named process, built on the stack you already use: Copilot Studio, Power Automate, Python or an MCP server.
- Guardrails: least-privilege tool access and a human approval step for any action that writes, sends or deletes.
- A test suite proving the automation does what it claims, and a short honesty report on what it does not catch.
- A before and after measure of time per task, taken on your real workload after the build, not estimated up front.
Not included
- Rebuilding a whole workflow. One process per sprint.
- Hosting and running it for you afterwards. That is Agent Care.
- A promised saving. The number comes from measuring after the build, and it can be small.
Proof of work
Tell me the one manual task eating your team's time, and I will tell you which part is actually safe to hand to an agent.
Start this conversation → -
03 / Services
Agent Care
From USD 600 / month · Monthly, cancel any time
Who this is for
Teams that finished a Review or a Build Sprint with me and want the boundaries kept intact as the agent changes.
The problem
Agents drift. A new tool, a model upgrade or a prompt change can quietly reopen a boundary that was closed at launch, and nobody re-tests it.
What you get
- A re-run of the review checks on releases you flag, up to two a month.
- Small fixes and rule updates, up to 6 hours a month.
- A one-page monthly note: what changed, what was tested, what is still open.
Not included
- Round-the-clock on-call or incident response. Replies come on working days.
- New features or new agents. Those are a new sprint.
Only offered after a Review or a Build Sprint, so we both know the system before I look after it.
Start this conversation → -
04 / Services
Security Technical Writing
Quoted per piece · Typically 1 to 2 weeks
Who this is for
Security or dev-tool vendors that need technical content developers actually trust, or engineering teams that want an internal security postmortem or advisory written up properly.
The problem
Most vendor security content is either thin marketing copy or too academic to act on. Developers trust writers who show their own tools failing and getting fixed, not writers who only show wins.
What you get
- One technical deep-dive: a CVE analysis, a benchmark write-up, or an incident postmortem, structured around evidence rather than adjectives.
- Code-level detail where it earns its place, and an explicit limitations section every time.
- Delivered ghostwritten or under your byline.
Not included
- Claims I cannot reproduce. If I cannot test it, it does not go in the piece.
- Pure marketing copy.
Proof of work
If your last security write-up got zero engagement from developers, send it over and I will tell you what is missing before you write the next one.
Start this conversation →