Attack research on prompt injection, MCP and agent attacks, agent kill chains, and hackbot design, plus findings from real engagements.
The word agent gets used for a chatbot with a system prompt, a script in a loop, and a system that plans and acts on its own. Three different animals. Only one can hurt you. A red teamer's guide to what an AI agent actually is and what agentic really means.
The autonomy that lets an agent clear your overnight queue is the same autonomy that lets one poisoned document turn it against you. One property, two sides. Why the more agentic the system, the further a single injection travels.
You will never make the model deterministic. So stop trying. You make the system around it deterministic, shrink the part that needs judgment to almost nothing, and put a fence around what's left. Here is how that actually works.
Prompt injection, jailbreaks, exfiltration, poisoning, tool abuse, excessive agency. The map of how every AI system breaks, whether you are attacking one or defending your own agent. A red teamer's working taxonomy.
Every team is wiring agents to MCP servers with the same casual trust they once showed Dropbox installs. The attackers have already noticed.
Claude Code is the best AI tool I've found for offensive security, but these concepts work with Gemini and other powerful AI agents too. Here's how to build your first AI red team skill and why it changes everything about how you operate.
A manual recon of an AI application takes 2-4 hours. With Claude Code and kali-mcp, it takes 12 minutes. Here's the exact workflow with working skills you can copy.
Most AI security tools test for prompt injection with a list of 50 known payloads. That's a scanner, not a pentester. Here's how to build an AI that actually thinks like an attacker.
Your company's AI chatbot trusts its vector database like gospel. I can change what it believes with a single document. Here's how RAG attacks work and how to test for them.
A jailbroken chatbot says something embarrassing. A jailbroken AI agent with database access and API keys does something catastrophic. Here's how to test agent security before an attacker does.
One AI agent found the vulnerability. A second built the exploit. A third exfiltrated the data. The whole operation took 4 minutes. Here's how to run coordinated AI red team operations.
The hardest part of any pentest isn't finding the vulnerabilities. It's explaining them to someone who doesn't know what a prompt injection is. Here's how AI generates better reports than most consultants.
