For copilots, chat, RAG, agents and MCP
Prompt injection, data leaking between customers, tool and MCP misuse, and what your agents are allowed to do. We test your AI feature the way attackers do, prove what we find, and sign a report your enterprise customers will accept.
What you get
For: SaaS teams shipping AI features to enterprise customers: an AI questionnaire to answer, an agent or MCP launch, or ISO 42001 to prepare for.
Granted at sign-in
The assistant's token is granted all 11 scopes it asks for at sign-in. After the fixes it holds 6.
agent activity · fernhill-crm-mcp
before testingafter fixes$ replay injected-note.json
Least-privilege score
11 scopes6 scopes45/100100/100
What we test
01
Direct, and indirect through the documents, web pages, emails and tickets your feature reads.
02
Whether one tenant's data can reach another through retrieval, memory, tools or logs.
03
What each tool and agent is allowed to do, and whether outside text can make it do more.
04
OAuth and token scopes, consent and tool permissions: the 32 published agent-access checks.
05
System-prompt and secret exposure, and what the model's output is allowed to trigger.
06
Findings mapped to OWASP LLM and Agentic Top 10 and ISO 42001, with answers for AI questionnaires.
Findings mapped to
OWASP Top 10 for LLM ApplicationsOWASP Top 10 for Agentic ApplicationsOWASP MCP Top 10MITRE ATLASISO/IEC 42001SOC 2ISO/IEC 27001
The 32 agent-access checks Free MCP server check Our MCP scopes study
How it runs
1
A free 30-minute call, then one fixed price in writing within a working day. Nothing is tested before you sign an authorization.
2
Our tools and AI agents map every endpoint, role and AI feature in scope, and flag what looks wrong. Fast and wide.
3
A tester confirms each lead by hand, tests what AI misses (authorization, tenants and business logic) and records the evidence.
4
A report and letter signed by Subcloak, a readout with your engineers, and a free retest once you've fixed things.
FAQ
Something else? Email hello@subcloak.com or book a call.
Do, it's a good start. Automated prompt suites find the obvious failures. What they don't give you is a test of your feature in its business context (whose data it can reach, what its tools can change), fixes that keep it useful, and a report your customers and auditors accept because someone independent signs it.
Yes. The 32 published agent-access checks cover OAuth and token scopes, consent and tool permissions on MCP servers and connectors. The free MCP server check shows a first view in seconds.
The OWASP Top 10 for LLM Applications, the OWASP Top 10 for Agentic Applications, the OWASP MCP Top 10 and MITRE ATLAS, with ISO/IEC 42001, SOC 2 and ISO 27001 items where they apply.
Yes. When the AI feature lives in the app we're already testing, it's $2,500 on top of the Proof Pentest.
Other tests
For SOC 2, ISO 27001 and enterprise deals
AI covers your whole web app and API in days. Then a person tests every role and tenant by hand, proves each finding, and signs the report your auditor and customers will accept.
$4,900–9,500≈€4,400–8,500≈£3,700–7,200≈CA$7,000–13,500≈A$7,000–13,600≈S$6,300–12,200≈AED 18,000–34,900from ₹95,000
Report in 5 working days
After a supply-chain attack
When the next npm, PyPI or AI-agent supply-chain attack hits, know within 48 hours whether it reached your developers' Macs, your CI or your secrets, then close the paths it used. Cloak Watch finds it; we fix it.
$4,500≈€4,000≈£3,400≈CA$6,400≈A$6,400≈S$5,800≈AED 16,500₹1,35,000flat
Answer within 48 hours
Hardening: $3,500–12,000≈€3,100–10,700≈£2,600–9,100≈CA$5,000–17,100≈A$5,000–17,200≈S$4,500–15,400≈AED 12,900–44,100₹1,05,000–3,60,000
Thirty minutes, no slides. Tell us what you need tested and when you need the report. You'll have one fixed price in writing within a working day.