> From: Andrew Daviel [mailto:advax@xxxxxxxxx] > Sent: Wednesday, July 07, 2004 2:30 PM > > A recent New Scientist article referred to the fact that > "sensitive data" may persist in computer memory, and be > swapped to disk and persist after a power-down. > > I had observed a while ago that text such as credit card numbers > entered into a form in Netscape could persist in RAM after the > application exits, and this seems to be still true for Mozilla. > > As discussed earlier in Bugtraq ("When scrubbing secrets in > memory doesn't work", 19 Nov 2002), in Linux/Unix the mlock() call can > be used to discourage swapping (MmLockPagableSectionByHandle ? in > Win32), while overwriting can be used to erase freed memory (as is done > in Gnupg). This conflates two related but different security issues: sensitive data left in RAM, and sensitive data left on disk after having been paged out. While the underlying problem (sensitive data outlasting the scope of its proper use) is the same, the parameters are quite different (for example, power-cycling generally takes care of the former but doesn't affect the latter), so they need to be analyzed and addressed separately. > It occurs to me that, while an unprivileged process cannot read system > memory directly, that a simple allocation of a large chunk of > memory might get data freed up or abandoned by previously running > processes. That shouldn't happen with any modern general-purpose OS. There's an Orange Book requirement called "object reuse" that requires all systems certified at C2 or above remove old data from an object (such as a region of memory) when it is reallocated to a new process. That's a simple guarantee for a virtual-memory OS to provide when allocating memory, so all the significant players do. > Given the now common practice of leaving computers powered on with > "high-speed" internet access, and the recent appearance of > trojans such as Bankhook.A and Pwsteal.Refest, I suggest that best > practice be updated to include the erasure and protection of "sensitive > data". Many people would say they already do include that. However, I think simply naming "best practices" is not particularly useful, in the long run. What you need is a weighted threat model, so you can address threats in an appropriate order. (The exact metric is debatable, but it should probably combine attack probability, likely degree of damage, and at a lesser weight the effort of implementing defense. And, of course, where it's trivial to protect against a threat, it's worth adding that protection even if the threat is unlikely.) For the general populace, I suspect "phishing" and other technically unsophisticated social-engineering attacks are likely more prominent than and equally damaging as the considerably more difficult class of attacks that involve gaining unfettered read access to physical memory and disk paging areas and trolling them for sensitive data. Of course, it's trivial to memset over a sensitive area when you're done with it, so programs ought to do so. Locking pages to prevent them from being written to disk may be more difficult: if it doesn't require special privilege then it's a potential DOS against physical memory resources, and if it does, then you may have to grant programs more privilege than they should have, creating a worse security hole. > One (probably very CPU-intensive, for some apps) way to enforce this > behaviour for malloc'd memory would be to make free() do an erase > operation as a system option. Creating "secure_free()" would > be better. I doubt you'd notice the CPU cost for nearly any normal application. However, as I noted above, the "object reuse" C2 requirement is widely satisfied and makes clearing memory when it's freed largely unnecessary. A much better approach is to have each program clear its own sensitive data as soon as possible. -- Michael Wojcik Principal Software Systems Developer, Micro Focus
