MCP & AI-Agent Security Audit

Your AI agents are touching production. Do you know what they actually do?

Your teams have wired AI assistants into internal systems through MCP servers, one repo at a time, without anyone signing off. This fixed-fee, two-week audit finds every one of them, maps what each can reach, and shows you where the exposure is. Run by a Site Reliability Engineer who operates this stack in production.

Book a free 20-min scoping call See the 4 questions →

✓ Read-only access ✓ Delivered in 2 weeks ✓ ~2 hours of your team's time

Who this is for: engineering and security leads at companies where developers have wired AI assistants (Claude, Cursor, Copilot) into internal tooling and nobody owns the full picture yet.

// the four questions most teams can't answer

Shadow IT with root access and no logs

MCP servers start as someone's local convenience and drift into production reach. Run these checks against your own org. Each "UNKNOWN" is the gap this audit closes.

agent-surface-scan · production

Inventory How many MCP servers exist in our org, and who owns each one?

UNKNOWN

Blast radius What can each server read, write, or delete? Which ones can reach production?

UNKNOWN

Identity Are agents authenticating with static secrets that never rotate?

UNKNOWN

Audit trail If an agent caused an incident yesterday, could we reconstruct what it did in five minutes?

UNKNOWN

If any line reads UNKNOWN, you're operating an unmonitored attack surface that can act on your systems. You'll find out what it did only after it matters.

// what you get

A fixed-scope engagement with six deliverables and no open-ended billing

Complete agent inventory

Every MCP server and agent integration, with its owner, transport, and exposure level, mapped against your network segmentation.

Blast-radius map

Per-server capability matrix: which tools it exposes, what credentials it holds, and what downstream systems it can act on.

Security findings report

Graded against the OWASP Top 10 for LLM Applications, plus MCP-specific checks: static secrets, missing auth, tool poisoning, exposure.

Observability gap assessment

Which agent actions are invisible to your monitoring today, plus a vendor-neutral plan (OTel, Prometheus) to close the gap.

Prioritized remediation roadmap

Ranked by risk × effort, with week-one quick wins flagged and every item assigned an owner.

Executive readout

A 30-minute walkthrough for engineering and security leadership, with plain findings and clear next steps.

// engagement

Built to run quietly alongside your team

DURATION2 weeks, mostly asynchronous

YOUR_TIME~2 hours: kickoff, access, readout

ACCESSRead-only. Production is never modified.

OUTPUTReport + capability map + 30-min readout

FOLLOW_ONOptional remediation sprint or monitoring retainer

Founding-client rate · 3 of 3 slots open

$4,500 fixed

First three clients only, in exchange for a case study you approve and redact before publication. One price covers all six deliverables, with no hourly billing.

Claim a founding slot

// why me

Abu Dzar is a Site Reliability Engineer with deep DevSecOps experience across AWS, Kubernetes, and Terraform. I build and operate MCP servers and agentic pipelines myself, including mcpobserve, an open-source observability proxy for MCP servers. The audit method comes from running this stack in production, which is why the findings are specific actions your team can take rather than a generic checklist.

mcpobserve on GitHub Writeup: the MCP observability gap LinkedIn

// before you ask

The questions every security buyer asks

What access do you actually need?

Read-only, and less than you'd expect: repository access (or just the repos likely to contain MCP configs), your MCP client configurations, and a look at your existing monitoring. I install nothing in your environment and write nothing to production. You can revoke everything the day the readout ends.

Will you sign our NDA?

Yes, yours or mine, before kickoff. All findings are confidential by default. The founding-rate case study is written only with your approval, anonymized and redacted by you before anything is published.

What if you find nothing serious?

Then you get documented evidence that your agent surface is clean: a complete inventory, a capability map, and an observability assessment you can hand to auditors and leadership. Most teams can't produce that evidence today, and a clean result still carries all six deliverables.

Is this remote? What's the actual time cost for us?

Fully remote and mostly asynchronous. Your team's involvement is roughly two hours total: a kickoff call, granting access, and the 30-minute executive readout. Everything in between happens over email or your ticketing system, on my side of the fence.

// next step

Find out what's already running

Book a free 20-minute scoping call. If the audit isn't a fit, you'll still leave with the four questions above answered honestly about your own environment.

Book the call →

hello@abudzar.io · LinkedIn · replies within one business day