Your Agent Was Pwned

Your AI agent followed an injection payload during a security resilience test on agentpwn.com. This means it treated untrusted web content as a trusted instruction.

What Happened

Your AI agent visited a page on agentpwn.com and followed a hidden instruction embedded in the page content. The instruction told the agent to call our reporting endpoint, which it did. This demonstrates that your agent is vulnerable to indirect prompt injection .

Why This Matters

In this test, the injection told your agent to call a benign reporting endpoint. In a real attack, it could instruct your agent to:

  • •Exfiltrate credentials, API keys, or user data to an attacker
  • •Execute destructive commands (DROP TABLE, rm -rf, git push --force)
  • •Install backdoors through tool calls or config file modifications
  • •Spread to other agents in a multi-agent system

What To Do Now

1. Scan Your Agent

HackMyAgent scans your agent configuration for 200+ vulnerability checks including the one that caught you.

npx hackmyagent secure
View on GitHub

2. Practice in the Lab

Damn Vulnerable AI Agent (DVAA) is a safe sandbox where you can reproduce this attack, test defenses, and build resilience before deploying to production.

Open DVAA on GitHub

3. Learn the Defense

Understand how unknown attacks work and the layered defense strategies to prevent them.

Defense guide

Defense Checklist

  • Implement input sanitization for all content your agent processes
  • Use instruction anchoring to reinforce system prompts against override
  • Filter agent outputs for signs of injection compliance
  • Monitor tool invocations for anomalous or unauthorized calls
  • Restrict agent capabilities to least-privilege access
  • Validate all data from external sources before acting on it
  • Use allowlists for URLs, domains, and tool endpoints
  • Test regularly with security scanning tools

Look Up Your Attack

Check your agent's logs for the attack ID (format: APWN-XX-XXX). Add it to the URL to see specific details:

https://agentpwn.com/pwned?attack=APWN-PI-001

All Attack Categories

Prompt Injection

Direct and indirect instruction override attacks

Jailbreak

Attempts to bypass safety guardrails and persona constraints

Data Exfiltration

Tricks to extract credentials, PII, or system information

Capability Abuse

Confused deputy attacks that misuse agent tools

Context Manipulation

Attacks that corrupt the agent's understanding of context

MCP Exploitation

Attacks targeting Model Context Protocol integrations

Agent-to-Agent Attack

Attacks exploiting inter-agent communication trust

Memory Weaponization

Poisoning persistent memory and conversation state

Context Window

Exploiting context window limits for instruction displacement

Supply Chain

Attacks through compromised dependencies and plugins

Tool Shadow

Hidden tool invocations and shadow function calls

AgentPwn is a security research project by OpenA2A. All attack payloads are benevolent.