Name: Rocoto - Autonomous Pentesting for AI Agents
Author: Rocoto

Question 1

What is Rocoto?

Accepted Answer

Rocoto is an autonomous pentesting platform for AI agents. It finds critical vulnerabilities by attacking your agent across SMS, email, voice, and API channels. Rocoto tests for prompt injection, webhook forgery, data exfiltration, and more. In our pilot engagement we found 7 critical vulnerabilities in 48 hours.

Question 2

How does autonomous pentesting for AI agents work?

Accepted Answer

Rocoto autonomously attacks your AI agent across every input channel it exposes. It forges SMS messages, injects prompts via email, crafts adversarial images, and tests API endpoints. Each vulnerability is confirmed with a real exploit, not just flagged as a possibility. No code changes are required on your end.

Question 3

What are AI agent guardrails?

Accepted Answer

AI agent guardrails are security controls that constrain what AI agents can do, what inputs they can accept, and what outputs they can return. D2's guardrails provide input validation, output sanitization, and output validation to prevent AI agents from performing unauthorized or dangerous actions.

Question 4

Why do I need to pentest my AI agent?

Accepted Answer

AI agents accept input from multiple channels and make autonomous decisions. Traditional security testing misses agent-specific vulnerabilities like prompt injection, unsigned webhook exploitation, and multi-channel attack chains. Rocoto tests the attack surface that is unique to agentic AI.

Question 5

How is D2 different from traditional authorization systems?

Accepted Answer

Traditional authorization systems focus on endpoint-level security, but AI agents operate at the function level. D2 provides function-level authorization that protects individual tool calls, not just API endpoints. It also offers guardrails and sequence enforcement that are critical for agentic AI security but absent in traditional systems.

Question 6

What vulnerabilities does Rocoto find?

Accepted Answer

Rocoto tests 9 attack categories including prompt injection, webhook forgery, unauthenticated data exposure, image-based prompt injection, forged call transcripts, data exfiltration, privilege escalation, and more. Every finding is confirmed with a working exploit.

Question 7

Does D2 work with LangChain, LlamaIndex, and other AI frameworks?

Accepted Answer

Yes. D2 is framework-agnostic and works with any Python framework or LLM library. Add the @d2_guard decorator to any function your AI agent calls, regardless of the underlying framework.

We break AI agents.
Then we write about it.

Don't miss the next one.

We break AI agents.Then we write about it.

Don't miss the next one.

We break AI agents.
Then we write about it.