[collingreene]

Here is our experience building and using program analysis as part of our product security efforts at facebook: https://engineering.fb.com/security/zoncolan/.

Its run in both self-service (output to developers), guided (output to product security oncall of security engineers) and used ad-hoc to power up manual security reviews. Depending on the accuracy of each rule and the impact of the pattern of security flaw the rule finds it is promoted to ultimately output to developers directly.

It finds about a third of the security vulns we unearth each year.

[lawnchair_larry]

That’s been my approach as well. An astonishingly large number of companies think they can buy an off the shelf static analysis tool and pipe the default output to developers. That’s counterproductive. A very small percentage of developers will understand the output, be able to assess the exploitability/severity, and care about fixing it. One might think you could then just have them take the “better safe than sorry” approach and fix everything, but FP rates for all of the commercial tools make that completely untenable. At the same time, you can’t expect to convince small teams of developers to model everything out and define sources/sinks using some obscure DSL that they have to learn. But, there are classes of issues that are extremely high impact, but only low accuracy static analysis rules can find the candidates. It’s that part in the middle that you don’t want to throw out, but you need security experts to vet. Other cases with high confidence checks are appropriate to short circuit straight to the devs, but it’s a bad first step.