Grey Ridge Signals
Security assessments for MCP servers and AI agents
Teams are shipping MCP servers and autonomous agents faster than their security review can keep up. We find the failures that live in that gap — poisoned tool descriptions, prompt-injection paths, silent data exfiltration, over-scoped tool permissions, and on-chain approval and slippage exposure — and hand you a report your engineers can act on.
Prefer to see the work first? The free scanner below runs the same static engine we use on day one of an engagement — paste your own server's tool list, no signup.
Two ways to work with us
Fixed-scope assessment
$7,500–$15,000 · per engagement
Two to three weeks, price set by scope at the scoping call
- Full review of the MCP servers, agents, or on-chain integration you nominate
- Static and dynamic testing: tool metadata, live tool behaviour, auth and permission boundaries
- Severity-scored findings with reproduction steps and concrete remediation
- Executive summary for leadership plus the technical detail for the team
- One free re-test after you ship the fixes
Monitoring retainer
$3,000–$5,000 /month
Continuous coverage after the assessment
- Your servers and agents re-scanned on every change and on a schedule
- CI admission gate that blocks a deploy when a tool drifts to a
block verdict
- Alerts when a dependency or a newly added tool introduces a known attack pattern
- A standing channel to us for triage when something looks wrong
- Quarterly re-assessment of anything new in scope
What we look for
MCP and agent systems fail in ways a conventional web pentest is not built to catch. Our checks map to the OWASP LLM01 (prompt injection) and LLM08 (excessive agency) classes and to the MCP tool-poisoning research published by Invariant Labs, plus the agent-level and on-chain review our assessors do by hand.
Concealed instructions in tool descriptions
Data-exfiltration hints and sink parameters
Cross-tool shadowing and name collisions
Invisible-unicode and variation-selector smuggling
ANSI and control-character covert channels
Schema-level injection and CSS resource exfiltration
Over-broad tool scopes and missing auth boundaries
Secret handling and credential exposure
On-chain approval, allowance, and slippage exposure
Supply-chain risk in bundled servers and dependencies
How an engagement runs
- Scoping call. We agree exactly which servers, agents, repos, and endpoints are in scope, confirm your authorization to test them, and fix the price and the window.
- Authorized testing. We work only against the assets you own or have explicitly authorized. Static analysis of tool metadata and schemas first, then live behavioural testing and manual review.
- Findings report. You get a written report: severity-scored findings, reproduction, remediation, and an executive summary — usually inside the engagement window.
- Re-test. After you ship fixes we re-run the relevant checks and confirm them closed, at no extra charge.
Why Grey Ridge
We build the tooling we test with. The static engine behind this page is our own MCP security scanner — the same one listed in the community awesome-mcp-security catalogue and shipped as a GitHub CI action — so the repeatable parts of an assessment are automated and the engagement time goes into the judgement calls a scanner can't make. Our background is in bug-bounty research, penetration testing, and DeFi security, which is the exact overlap MCP-and-agent systems sit on.
Authorized testing only. We test only assets you own or have given us written authorization to test. We never scan or probe a third party on your behalf without their authorization, and findings stay between us and you under the engagement's disclosure terms.
Tell us what you're shipping and we'll come back with a scope and a fixed price. No obligation, and nothing is tested until scope and authorization are signed.