Microsoft called its first month of predicting whether hackers will create exploit code for its bugs a success - even though the company got its forecast right less than half the time.
"I think we did really well," said Mike Reavey, group manager for the Microsoft Security Research Centre (MSRC), when asked for a post-mortem evaluation of the first cycle of the team's Exploitability Index.
"Four of the issues that we said where consistent exploit code was likely did have exploit code appear over the first two weeks. And another key was that in no case did we rate something too low."
Last month, Microsoft launched the index , which rates each vulnerability using a three-step system that, in descending order of severity, said researchers or hackers would come up with a consistently working exploit, develop an exploit that worked only some of the time, or fail to craft attack code at all.
The predictions were valid for the following 30 days, or until the next cycle of patches was released.
Of the nine October vulnerabilities marked "Consistent exploit code likely," four did, in fact, end up with exploit code available, said Reavey, for an accuracy rate of 44%. None of the nine tagged "Inconsistent exploit code likely" had seen actual attack code. But Microsoft correctly called the four bugs last month tagged with the label "Functioning exploit code unlikely." As Reavey said, exploit code did not appear for any of the four.
All told, Microsoft correctly predicted eight out of October's 20 vulnerabilities' exploitability, an accuracy rate of 40%. (One of the month's 21 bugs did not receive a rating, as Microsoft said public exploit code was already circulating, making a label moot.)
That accuracy rate was down slightly from what Microsoft claimed during a five-month internal run of the index before it announced the program in August at the Black Hat security conference. According to a presentation Reavey gave at the conference, during that the five months it assigned ratings, Microsoft correctly predicted the exploit code availability of 17 out of 36 bugs, for an accuracy rate of 47%.