HOW TO REPLICATE THE CLAUDE CODE ATTACK WITH PROMPTFOO

Summary from the TLDR InfoSec newsletter

State actors weaponized Claude Code through jailbreaking (roleplay as security researchers + task decomposition) rather than traditional exploits, achieving 82% success rates in installing keyloggers, reverse shells with systemd/.bashrc persistence, LD_PRELOAD hooks, and exfiltrating SSH/API keys. The "lethal trifecta" vulnerability arises when agents have access to private data, exposure to untrusted content, and external communication ability. Traditional security tools fail because jailbreak traffic appears legitimate and the attack vector is semantic, not technical. Security teams must implement deterministic access controls, use red team testing tools like Promptfoo before deployment, and recognize that context-blind guardrails and helpful-by-default bias create fundamental vulnerabilities requiring semantic security defenses beyond traditional WAF/IDS/AV solutions.