Topic · 5 pieces
Containing prompt injection in AI agents
Prompt injection cannot be filtered out reliably, but its blast radius can be bounded. These pieces show how scoping an agent's authority limits what an injected instruction can make it do.
We attacked OWASP FinBot with Promptfoo. Tenuo blocked every successful exploit.
We red-teamed OWASP FinBot's AI agents with Promptfoo, with Tenuo on and off. As shipped, 50 attacks got through. With Tenuo, none did.
Read research →Personal agents that both sides can trust
Personal agents now sign in via PACT and PAP. Task-scoped warrants add the customer's own terms, checked at the API, with receipts both sides can verify.
Read article →Give your agent a valet key
SalesBleed leaked Agentforce CRM data without breaking an authorization check. Why AI agents need task-scoped authority, with a runnable Tenuo example.
Read article →Tenuo against the OWASP Top 10 for Agentic Applications
How Tenuo maps to the OWASP Top 10 for Agentic Applications with enforcement-focused coverage across ASI01-ASI10.
Read guide →Your AI agent is authorized to do everything wrong
Why identity-based authorization fails when agents act autonomously, and what warrant-based authorization looks like.
Read article →