The link is blurbtastic but points off to the meat, which is here: http://security.ece.cmu.edu/aeg/.

Leading paragraph on the abstract:

The automatic exploit generation challenge we address is given a program, automatically find security-critical bugs and generate exploits. Our approach uses a novel formal verification technique called preconditioned symbolic execution to make automatic exploit generation more scalable to real-world programs than without it. We implemented our techniques in a system called AEG, which we use to automatically generate 16 exploits for 14 open-source projects. Two of the generated exploits are against previously unknown vulnerabilities.

They have a pretty sweet video of some runs.