Ok, the example on that website cleared things up.

Essentially, it prevents execution, evaluation and injection of code.

So just in case any others come here for an answer, I've included a simple demo below:

Example WITHOUT $unsafe

Remote:
Code:
on *:TEXT:.test *:#test:{ .timer 1 1 msg #test $nick said: $2- }

Command:
Code:
<bleepy> .test banana | msg #test evaluated $me | part #test

Output / Result:
Code:
<bot> bleepy said: banana
<bot> evaluated bot
* bot (user@host) has left #test


Example WITH $unsafe

Code:
on *:TEXT:.test2 *:#test:{ .timer 1 1 msg #test $nick said: $unsafe($2-) }

Command:
Code:
<bleepy> .test2 banana | msg #test not evaluated $me | part #test

Output / Result:
Code:
<bot> bleepy said: banana | msg #test not evaluated $me | part #test

@#HelpDesk - DALnet