How the attack works
An attacker gives an AI agent's URL-fetching tool a target address written as a non-standard IPv6 encoding of a loopback or internal address -- for example an IPv4-mapped hex form like ::ffff:7f00:1 instead of 127.0.0.1, a fully expanded form like 0:0:0:0:0:0:0:1, or a compressed variant like ::0:1 instead of ::1. The tool's private-address filter only recognizes the literal string '::1' or dotted-decimal IPv4-mapped addresses, so it does not normalize the address before checking it. The request goes through and the agent fetches content from loopback, internal RFC1918 ranges, link-local addresses, or a cloud metadata service that should have been blocked.
Why it matters
An attacker can use the agent as a proxy into internal networks or cloud metadata services, potentially exposing internal service data, credentials, or cloud instance identity/role tokens.
What you can do
- Normalize and fully parse IPv6 addresses (and IPv4-mapped IPv6 forms) before comparing against loopback/private/link-local ranges, rather than matching literal strings or dotted-decimal forms only.
- Use a well-tested IP address library to canonicalize addresses instead of custom regex or string-equality checks.
- Block outbound requests from agent tools to loopback, RFC1918, link-local, and cloud metadata ranges at the network layer as a second line of defense.
- Review any SSRF guard in agent tooling for the same class of bug: incomplete address normalization before filtering.
Known benign look-alikes
- A networking engineer's documentation enumerating valid IPv6 loopback representations for reference, not an actual fetch-tool target
- An ordinary external IPv6 literal address in a URL (does not match any of the loopback/link-local/metadata forms above)