There were rumored to be "explosive bolts" on the ARPA/MILNET gateways (whether they were metaphorical or not, I don't know).

Here's something interesting that Milo Medin wrote about dual homed sites like NSA and NASA, that were on both the ARPANET and MILNET:

    To: fair@ucbarpa.berkeley.edu (Erik E. Fair)
    Cc: Hackers_Guild@ucbvax.berkeley.edu, ucdavis!ccohesh@ucbvax.berkeley.edu
    Subject: Re: a question of definition
    Date: Thu, 29 Jan 87 15:33:35 PST
    From: Milo S. Medin (NASA ARC Code ED) <medin@orion.arpa>

    Right, the core has many gateways on it now, maybe 20-30.  All the LSI's will
    be stubbed off the core however, and only buttergates will be left after
    the mailbridges and EGP peers are all converted.  Actually, I think DARPA is
    paying for it all...

    Ames is *not* getting a mailbridge.  You are right of course, that we could
    use 2 gateways, not just 1 (actually, there will be a prime and backup anyways),
    and then push routing info appropriately.  But that's anything but simple.
    Firstly, the hosts have to know which gateway to send a packet to a given
    network, and thus have to pick between the 2.  That's a bad idea.
    It also means that I have to pass all EGP learned info around on the
    local cable, and if I do that, then I can't have routing info from
    the local cable pass out via EGP.  At least not without violating
    the current EGP spec.   Think about it.  It'd be really simple to
    create a loop that way.  Thus, in order to maximize the use of both
    PSN's, you really need one gateway wired to both PSN's, and just
    have it advertise a default route inside.  Or use a reasonble IGP,
    of which RIP (aka /etc/routed stuff) is not.  I'm hoping to get
    an RFC out of BBN at this IETF meeting which may go a long way in
    reducing the use of RIP as an IGP.

    BTW, NSA is an example of a site on both MILNET and ARPANET but without
    a mailbridge...

    There is no restriction that a network can only be on ARPANET or MILNET.
    That goes against the Internet model of doing things.  Our local
    NASA gatewayed nets will be advertised on both sides.  The restriction
    on BARRNet is that the constituent elements of BARRNet do not all
    have access to MILNET.  NSF has an understanding with DARPA and
    DCA that NSFnet'd sites can use ARPANET.  That does not extend to
    the MILNET.  Thus, Davis can use UCB's or Stanford's, our even NASA's
    ARPANET gateways, with the approval of the site of course, but
    not MILNET, even though NASA has MILNET coverage.  Thus we are required
    to restrict BARRNet routing through our MILNET PSN.  If we were willing
    to sponsor UCB's MILNET access, for some requirement which NASA
    had to implement, then we would turn that on.  But BARRNet itself will
    but cutoff to MILNET (and probably ARPANET too) at Ames, but not
    cut off to other NASA centers or sites that NASA connects.  There is
    no technical reason that prevents this, in fact, we have to take
    special measures to prevent it.  But those are the rules.  Anyways,
    mailbridge performance should improve after the conversion, so
    UCB should be in better shape.  And you'll certainly be able to
    talk to us via BARRNnet...  I have noticed recently that MILNET<->
    ARPANET performance has been particularly poor...  Sigh.

    The DCA folks feel that in case of an emergency they may be
    forced to use an unsecure network to pass certain info around.  The
    DDN brochure mentions SIOP related data for example.  Who knows,
    if the balloon goes up, the launch order might pass through Evans
    Hall on its way out to SAC...  :-)


                        Milo