It depends on the
rand() implementation... the standard (POSIX / Single Unix) gives a sample implementation but any system is free to have something better.
You can see the sample code there. If that exact code is used, then the internal state is a 32-bit integer, and it suffices to get two successive 16-bit output values to recompute the internal state. Actually, any 32 bits of output are sufficient, with the generic reconstruction algorithm known as "brute force": it won't take long, for a computer, to try all possible 32-bit internal states until one is found, which matches the observed output. It is possible to recompute the state waaaaaaaaay faster by doing some linear algebra, but since brute force works well, why bother ?
In usual Linux systems,
rand() is an alias for
random() which is much better, but still not "good", as far as randomness goes. The
srandom() function still initializes the internal state with a 32-bit seed, which is amenable to brute force. You just lose the linear algebra shortcuts; or, at least, things require a bit more mathematics.
random() is not a cryptographically secure PRNG and its internal state can be rebuilt by observing some output bits. The simplest method is still brute force, and it works well.
(Also, if the application did not bother to call
srandom(), then the initial seed is always 1, so brute force will work very well.)