AI Agents
AI pentest agent for OWASP LLM Top 10 security testing
A multi-integration Azure AI Foundry agent connected to Xero, Outlook, SharePoint, a helpdesk and web search, built as a realistic testbed for prompt injection, data leakage and excessive-agency testing.
- Client
- 24by7 Security
- Industry
- Cybersecurity
- Status
- Delivered
The problem
The security team needed a realistic, connected AI agent to test how enterprise agents fail, and a documented mitigation plan they could act on.
What we built
- GPT-4.1 agent on Azure AI Foundry behind a Flask app on Azure Container Apps
- Integrations: Xero (via MCP server), Outlook and SharePoint (Graph, OAuth 2.0), Bing grounding and an IT helpdesk
- Test coverage across all 10 OWASP LLM risk categories
Challenges we solved
- Indirect prompt injection through emails and documents
- Cross-system data correlation
- Token and session security
Results
- Risk-rated catalogue of attack vectors
- Prioritised mitigation roadmap: RBAC, PII masking, audit logging, least privilege, monitoring
Technology stack
- Azure AI Foundry
- GPT-4.1
- Azure Container Apps
- Flask
- Xero MCP
- Microsoft Graph
- SQLite
Need something similar?
Free 30-minute scoping call. We reply within 24 hours.
Talk to us about AI Agents