Andy Davidson <andy@nosignal.org> writes:
Because TCP MD5 packets touch a router's CPU, using MD5 introduces a new attack vector - see nanogii passim (e.g. http://www.nanog.org/meetings/nanog39/presentations/Scholl.pdf). Don't do it. :-)
Tom's slide deck is often misinterpreted - the salient parts are on pages 13 and 16. The big problem is that random packets can hit the control plane - AT ALL. One can kill it just as easily with a synflood on some other tcp port (perhaps ssh or even one that it isn't listening on?). Hopefully your modern exchange point router has some sort of control plane policing. Indeed, hopefully your backbone routers have some sort of control plane policing mechanism and you have turned it on and subjected your policy to some scrutiny. Blowing up un-password-protected BGP sessions by spoofing has not turned out to be a big problem in recent years. It probably helps that the dangers of highly predictable ISNs have become well-known (and hopefully acted upon by router vendors but history has shown that making blanket statements about that sort of thing on NANOG is unwise). A read of http://tools.ietf.org/html/rfc6528 may prove instructional. Turning on or off md5 makes minimal difference in CPU loading either during an attack or not, but it is another thing to get wrong - operational complexity without significant reward. I agree with Andy's conclusion. Don't do it unless whoever you're peering with demands it. It's not worth the complexity to set it up in the first place, and it's not worth your time to argue against it if someone is quite convinced that enabling md5 on your bgp session will save the world. -r