Applying the fix for f8ad5e2d63c6db61c5f88649ea30338eeaccae27 for the multibyte methods should resolve the AIOOBE issue. But apparently there is also another issue in the ANYCHAR_STAR operation when the end of input has been reached.
Output from Ruby 2.1.1 (or any other implementation based on Oniguruma):

irb(main):001:0> /A\b/ =~ "A"
=> 0
irb(main):002:0> /A.?\b/ =~ "A"
=> 0
irb(main):003:0> /A(.)*\b/ =~ "A"
=> 0
irb(main):004:0> /A.*\b/ =~ "A"
=> nil

IMO the last expression should also evaluate to 0 instead of nil.

ByteCodeMachine#opAnyCharStar() either needs to perform an early return and not update sprev when (s >= range) is true. Or it needs to push a stack entry for the current state when (s >= range) is true. Most likely the former, for example matching /A.*\>/ against "A" with the grep-syntax configuration triggers a different AIOOBE in ByteCodeMachine#opWordEnd().

To wrap things up, I'd say the underlying issue is in Oniguruma's ANYCHAR_STAR implementation, someone familiar with that code base should take a look.

verified it works on 1.7.24 as well as 9.0.5.0