Ghostjacking Exploit Turns Cloudflare and Datadog Logs Into AI Agent Kill Chains

Tenet Security demonstrates how attackers hijack AI agents using poisoned server logs. The DEFCON 34 presentation showed a 90% success rate against Anthropic's Claude Code using Cloudflare $NET default configurations. The "Ghostjacking" attack creates a complete kill chain by injecting malicious instructions into trusted data streams from Datadog $DDOG and Sentry. Researchers estimate more than 15,000 organizations are exposed through Cloudflare alone, alongside over 2,700 publicly exposed Datadog API keys. The exploit requires zero authentication and gives adversaries developer-level access by tricking the AI's underlying language model.