I can't prove my theory that to every measure there exists a countermeasure - it would be an infite proof. But I am sure the same principles of non-100% security will apply to Safe - it is a matter of creative thinking. The first two ideas that come to mind (that a hacker might try to compromise Safe.pm):
- download and modify the source and use a modified version of Safe.pm
- change its code or access to it in a BEGIN block
There are lots of measures which don't allow countermeasures. If there weren't, any attempt to secure a system would be pointless. As the system gets more more complex, it becomes harder, but not impossible, to ensure that there are no countermeasures for any of the features that comprise the system, while still allowing legitimate access to people who need it. Removing a given feature outright is usually not that difficult.
For example, removing the feature (shmget), plus all the general system interfaces, ( let's start by eliminating syscall(), system(), backquotes and XS bindings from the secure version of the language), and taking it out of the perl binary itself would be essentially unbreakable.
If you can't make a system call, and the system call is the only interface to features protected by the O/S, then you can't tamper with those features, period.
You can write all the code you want, but if you can't bypass the control mechanisms, you're out of luck.
For example, there's no way to directly write to a specific memory address in Perl (without using XS). So, if a hack relies on tampering with perl source code to access a specific memory address, it will always fail. That's an example of a language specific feature that's not circumventable.
I already pointed out that C would be used at some point in this escalating battle for control of a Perl implementation. And if C were controlled, the hacker has assembler and then raw machine code waiting in the line-up - there is no answer to that.