For the SGML bit, rather than using <[^>]*?> it would be better to have <[^>]+> to avoid matching empty elements like <> Crossing fingers that this time it will get through the markup engine ;-)

Couple of things though:
1) US Phone Numbers
IANAA (I am not an American…) If isolating the expression so that it doesn’t feature as part of a longer sequence, why not just put \b at the start and end of the regex so that it has to have a word boundary each end? Assuming that neither area codes nor exchange codes can start with a zero, you could restrict it a little more with
[A-PR-Y1-9][A-PR-Y0-9]{2}-[A-PR-Y1-9][A-PR-Y0-9]{2}-[A-PR-Y0-9]{4}

2) Non-entity ampersands
You could probably tighten it up a little, by using [a-z] instead of \w - I don’t know for certain, but I *think* named entities are always a-z chars, not underscores, digits, etc. Also, it might be possible to restrict the number of digits in the number entity version. Overall, though, it’s a classic example of why parsing HTML with regexes is a Bad Thing(tm)

3) Floating point numbers
You could avoid the problem of passing things like 123.4567e4 by just removign the first \d* from the expression - maybe that’s what people meant when they said it’s a bug? Also, what was the reason for capturing the part that comes before the decimal point? Leaving non-capturing groups aside, what’s the benefit of putting this part in brackets?

4) MAC addresses
If you disregard the other forms for a MAC address, one severe failing of this regex is that the {5} quantifier means that the regex will only match MAC addresses that have the same octet five times, e.g. AB:AB:AB:AB:AB:CD - maybe it would have been better to do something like
([\da-f]{2}:[\da-f]{2}:[\da-f]{2}:[\da-f]{2}:[\da-f]{2}:[\da-f]{2})
or
(([\da-f]{2}):([\da-f]{2}):([\da-f]{2}):([\da-f]{2}):([\da-f]{2}))
if you want to capture each octet separately. There’s also the POSIX character class [:xdigit:] for hex digits, but some people find [\da-f] easier to read!

5) SGML elements
]*?> will also match an empty element, like , since * will also match zero occurences - maybe better to have ]+> … You don’t need the non-greedy ? as it won’t change what the regex matches. You could look at restricting things further, maybe elements using a-z, etc, although again there’s the trying-to-validate-sgml-with-regex problem again!