How the attack works
A plugin (documented here as matching the FastGPT web-crawling plugin) accepts a request with a URL-shaped parameter such as url, target_url, fetch_url, callback, or redirect_uri. The attacker sets that parameter to an internal address instead of a public one: a loopback address, an RFC1918 intranet IP, a link-local or CGNAT address, or a cloud metadata service address. The plugin fetches that address on the attacker's behalf, without verifying it is external. This lets the attacker probe or pull data from internal-only services the attacker could not otherwise reach directly.
Why it matters
An attacker can use the AI agent's server as a stepping stone into internal infrastructure, potentially reaching admin panels, cloud credentials via metadata endpoints, or other internal-only systems, leading to information disclosure or further compromise.
What you can do
- Validate and restrict any URL-fetching plugin so it rejects loopback, RFC1918, link-local, CGNAT, cloud-metadata, and in-cluster (.svc.cluster.local) destinations by default.
- If internal fetches are a legitimate use case (e.g. crawling an internal wiki), maintain an explicit allowlist of approved internal hosts rather than allowing arbitrary user-supplied targets.
- Block outbound requests from agent/plugin infrastructure to the cloud metadata address and other sensitive internal endpoints at the network layer, independent of the plugin's own logic.
- Review logs for fetch/crawl requests where the URL parameter resolves to an internal address, and separately track expected exceptions like known dev/CI or internal-docs use to reduce noise.
Known benign look-alikes
- Agent legitimately asked to crawl or summarise an internal wiki, Confluence, Jira or documentation site that is hosted on an RFC1918 address and passed as a url= parameter.
- Local development and CI - a developer exercising the crawl/fetch tool against a service running on 127.0.0.1 or 0.0.0.0 with an explicit port.
- Kubernetes-hosted agents fetching an in-cluster API via a *.svc.cluster.local URL passed to a generic fetch tool.
- Tailscale / CGNAT estates where 100.64.0.0/10 addresses are the normal way to reach internal services.
- Authorised SSRF regression suites and scanner payload corpora replayed through the agent (partly suppressed by filter_vuln_docs).
- Configuration or seed data blobs that embed example private-range URLs (suppressed by filter_cidr_doc and filter_template_placeholder).