Attacking AI Agents
The attack surface of autonomous agents — prompt-injection chains, tool abuse, confused-deputy and exfiltration, with reproducible proofs of concept.
The AI Agent Attack Surface: A Practical Threat Model
A practical threat model of autonomous AI agent security: prompt injection, tool abuse, memory poisoning, excessive agency, and multi-agent propagation — mapped to the OWASP Top 10 for Agentic Applications and MITRE ATLAS.
Read →Indirect Prompt Injection in AI Agents: Attack Chains, Payloads, and Injection Techniques
How attackers chain indirect prompt injection to hijack AI agents: injection sources, payload construction, multi-step attack chains, and the techniques that carry hidden instructions into the model.
Read →Memory and Context Poisoning in AI Agents: Attack Patterns and Injection Techniques
How attackers poison an AI agent's long-term memory and retrieved context: injection patterns, persistence techniques, and how false data is planted to hijack every future session.
Read →Multi-Agent and Agent-to-Agent Attacks: How One Compromised Agent Poisons the Swarm
How a single compromised agent spreads malicious instructions through a multi-agent system, which trust assumptions enable lateral movement, and what isolation defenses can contain the blast radius.
Read →Tool and Function-Call Abuse in AI Agents: The Confused-Deputy Problem Explained
How attackers exploit AI agents' privileged tool access through the confused-deputy problem — mechanism, concrete example, and least-privilege defenses that actually work.
Read →