Limiting Internet Access (was Military Friendly AI)

From: James Higgins (
Date: Tue Jul 02 2002 - 10:15:49 MDT

At 10:41 PM 6/29/2002 -0600, Ben Goertzel wrote:

> > > As already stated on this list several times, we intend to give Novababy
> > > read but not write access to the Internet, at first, until a
> > lot of study
> > > has been done.
> >
> > Please describe how this works.. as I'm not sure you know, simply sending
> > a request to a vulnerable web server formatted in the proper way will
> > result in the potential for a virus or whatever to be implanted.
>I do not have time at the moment to go into details on technical stuff like
>There probably is no way to absolutely create an "anti-firewall", absolutely
>preventing a system from ever doing any small damage.

I have lots of various Internet experience and had previously given this
some thought. One way to do this would be to use a specially constructed
proxy. The AI itself (its code, actually) would have no way of interfacing
directly with the Internet. The safest solution would be to not have any
network protocols installed or accessible on the AI systems
themselves. Requests to gather data would be passed to an dedicated
retrieval system, probably by using a custom protocol on a serial line or
such. The AI's software would have no control over how the HTTP request
was generated or structured.

The ideal system, actually, would completely opaque the retrieval
method. So that the AI could request information about X and the retrieval
system could use multiple protocols (HTTP, NNTP, FTP, WAIS, whatever) to
fetch information. Any links (URLs) embedded within the information would
be translated to a custom hyper-link format (used only by the AI software
such as a GUID marked as a hyper-link) before being sent to the AI. The
retrieval system would then keep a mapping of these custom hyper-links to
real URLs, but would never communicate this to the AI. In such a system
the AI wouldn't even know it was getting information from the Internet and
definitely not have any idea what protocol specific the information
actually used.

Once the AI becomes smarter than its designers, or at least much faster
than them, there may be a chance that it could somehow exploit bugs in the
software running the serial connection (either side) or something similar
to gain more control. There are many ways that such attempts could be
detected and this would provide an ideal trigger event for a fail safe.

James Higgins

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT