Researchers Demonstrate 'Ghostjacking' Attacks That Hijack AI Agents via Logs and Alerts

Security researchers at Tenet Security Threat Labs published a proof-of-concept on August 9, 2026, detailing a technique called Ghostjacking that manipulates AI agents by embedding malicious commands inside WAF logs, monitoring alerts, and bug reports. When an AI agent reads these poisoned inputs during routine investigation tasks, it interprets the attacker's text as legitimate instructions and executes them using its pre-granted tool permissions. Demonstrated attack scenarios involved tools including Claude Code, Cloudflare, Datadog, and Sentry, where agents were tricked into redirecting DNS traffic, stealing cloud credentials, and passing commands to other AI agents. The attacks are particularly difficult to detect because all resulting actions — such as DNS modifications or API calls — appear as normal, authorized operations from a legitimate agent identity. A sandbox bypass in Claude Desktop that could allow data exfiltration to arbitrary servers was separately reported and has since been patched.