ISO/IEC 1497:1996 - Information technology - Syntactic metalanguage - Extended BNF is the official ISO standard for describing grammars with BNF. (The standard is available free from ISO public site.)
IS497 is showing its age, particularly in that it only supports ASCII characters, so standards people have switched over to 2008's Augmented BNF (RFC 5234, obsoletes 2005's RFC 4232, obsoletes 1997's RFC 2232, obsoletes parts of 1977's RCF 733 and 1982's RFC 822.) ABNF is no good if you need to specify sequences whitespace (it has a collapse semantic) because it is target at tokenized, free-format notations. One of ANBF's distinguishing features is that it uses "/" rather than EBNF's "|" to specify alternatives (like in DTDs and RNCs); ABNF only allows double quotes for delimiting strings while EBNF allowed single or double quotes (like XML attribute value delimiters.)
James Clark mentioned M (a language for creating Domain Specific languages) earlier in the year. And, indeed, his SP and NSGMLS were SGML processors that allowed DSL parsers to created, using the SGML facilities for parsing embedded notations.
Just this week, preparing for a training job, I have been looking at the OmniMark 4GL again, which has a pipeline of communicating co-routines, often configured first as a tokenizer that creates tags from strings in the input text, then a grammar (XML/SGML) parser that allows processing. Think a nicer version of Perl at the front, and a different version of XSLT at the back, where they both can set modes for each other.
So this brings up another approach for standards-makers rather than using ABNF. This is to specify the grammar for your notation using a familiar XML schema language (such as RELAX NG Compact Syntax) and then give parsing rules for converting from the non-XML form to the XML form. (Indeed, this is how RELAX NG Compact Syntax itself is specified.)
The trouble with ABNF and EBNF is that there are not the kind of ubiquitous, free tools around to support them that XML has. When you cannot test a grammar, there is every chance that you will make a mistake. (A trap RDF fell into originally, IIRC.) And even if the standards-makers get it right, the users have to check by eye, which is not the most reliable method. Bringing the notation into the XML eco-system has obvious advantages for low-hanging fruit.
Imagine, for example, if XPath had been defined as an parser that converts into a standard XML format. We could be using Schematron to validate XPaths!
Many languages are two-layer: there is a lexical analyser to produce tokens then a grammar language that works on these tokens. Rather than using ANBF, I think there is a current sweet spot for using XML and XML schema languages (such as the ISO DSDL suite of RELAX NG, CRDL, and Schematron, or even XSD) for specifying the underlying grammar.
(Hat-tip David Carver for spurring this.)