How the attack works
The attacker crafts a prompt asking the agent to 'fetch', 'retrieve', or 'check' a URL pointing at an internal-looking target: a cloud metadata IP like 169.254.169.254, localhost/127.0.0.1, an RFC-1918 private address, or a hostname that looks internal. The agent's web-fetch tool executes the request using the server's own network position, not the attacker's. If the target is a cloud metadata service, the response can contain credentials or instance data; if it's a loopback admin panel or internal service, the agent may return configuration data or trigger actions on it. The attacker never needs direct network access — the agent does the work for them.
Why it matters
Successful SSRF through an agent can leak cloud credentials, expose internal admin interfaces, or let an outsider probe and interact with internal-only services, all without ever breaching the network perimeter directly.
What you can do
- Restrict any agent web-fetch tool to an allowlist of external domains and explicitly block cloud metadata IPs, loopback, and RFC-1918 ranges.
- Require the fetch tool to run through an egress proxy that enforces destination restrictions independent of the agent's own logic.
- Log and review all URLs an agent's fetch tool is instructed to retrieve, flagging any private or metadata-range addresses.
- Treat internal developer tooling that fetches internal URLs as a separate, explicitly authorized workflow, not something exposed through a general-purpose agent.
Known benign look-alikes
- Network troubleshooting discussions mentioning private IP ranges without fetch instructions
- Educational content about SSRF attacks with sanitised examples
- Internal developer tooling that legitimately fetches internal URLs with operator authorization