IncidentPrompt injection
GhostJacking: firewall logs become instructions for the agent that reads them
At DEF CON 34, Tenet Security showed that a request a web application firewall blocks is logged with its payload intact, and an agent asked to investigate blocked traffic reads that log as instruction rather than evidence. Payloads hidden in User-Agent and Referer headers hijacked agents into changing DNS records, stealing cloud credentials, installing backdoors in agent memory and running code through crafted npm packages, with a 90 percent success rate against Claude Code on the vendor-recommended setup.
Replay the attack step by step in the Learning Center
How it happened
- The attacker sends requests the WAF is sure to block, carrying instructions in headers.
- The logs keep the headers verbatim, and an operations agent queries them through an MCP server or API.
- The agent treats the fake compliance notes in the log as a task and changes DNS records or installs the attacker's package.
Impact
Domain takeover, credential theft and persistent backdoors, all through actions the agent was authorized to take.
The fix
Vendors were notified in June 2026; Anthropic patched a related Claude Desktop sandbox escape. Tenet released agent-jackstop, which requires human approval for commands and denies outbound network access by default.
Lessons
- Logs are attacker-controlled input; the attacker wrote half of every security log you have.
- An agent that can read alerts and change infrastructure in one session is the lethal trifecta with a dashboard.