Hi there,

I'm building a site that needs to securely obfuscate attachments in a public S3 bucket. The solution I arrived at seems like it might be useful to the community as a general-purpose tool that remains customizable enough to allow developers with different requirements to implement it differently without having to hack into Paperclip core. It is a set of storage-agnostic extensions to Attachment and Interpolations and supporting tests.

It features:

• No extra data model required
• Choice of HMAC digest algorithm on a per-attachment basis via openssl (modeled after ActiveSupport::MessageVerifier, default algo is SHA1)
• Choice of fields to include in the hash on a per-attachment basis, which permits many developer-selected tradeoffs between security and flexibility -- some examples:
  • :hash_data => ":class/:attachment/:id" - Unique per asset, long-lived filename with deducible style URLs (e.g. you want a 3rd party site to be able to generate style URLs from a single URL trunk that you provide once and always get the latest attachment)
  • :hash_data => ":class/:attachment/:id/:style" - Adds non-deducible style URLs (e.g. you want to store uncompressed :originals but don't want end users to be able to find them and kill your bandwidth bill)
  • :hash_data => ":class/:attachment/:id/:style/:updated_at" - Adds non-deducible versions (e.g. if a user's private profile photo URL makes it out into the wild via a malicious friend, they can unfriend, upload a new photo and the leak is plugged) -- this is the default

Since the hashes are extremely unlikely to collide, you can also remove other elements of the :path that were previously required for global uniqueness (like :id), which can give your users plausible deniability of ownership, or even completely obliterate the directory structure (:path => ":hash" -- or less controversially :path => ":hash.:extension") such that an attacker can't even prove what kind of attachment he/she has gained access to. Paired with a large :hash_secret and HTTPS, your users are left with a similar security guarantee to the analog hole -- attackers would need an authorized-user mole, or would need to compromise either your server-side secret or an authorized user's machine in order to gain access to protected attachments, and then would be able to freely distribute them (sadly, using your hosting infrastructure as an accomplice).

The :timestamp bug

While building this, I ran across what I think is a bug: The :timestamp interpolation is non-deterministic in the presence of per-thread time_zones, e.g. displaying time zones per user location, which would affect any such site that uses :timestamp in an attachment :path or :url (or if I had used :timestamp in :hash). While using :timestamp in paths and URLs is probably a little-used feature given the general wordiness of Time#to_s, and the intersection of sites that attempt to use it as well as implement per-user time zones is a vanishingly small slice of Paperclip users, I think it's worth taking a look at.

I took a stab at dealing with it by adding a new attachment config parameter :use_default_time_zone (defaulted to true) that explicitly churns out :timestamp interpolations in the server-wide default time zone. This is not perfect, because it's possible that a site owner would want to change the default time zone via config.time_zone=, which would in turn invalidate all of their existing attachment URLs. One option available to such an implementor would be to redefine my new method Attachment#time_zone explicitly to meet their requirements. Or I could add a :time_zone option on Attachment, defaulted to null, and falling through to the present Attachment#time_zone implementation. Such an option could even optionally take a block a la options[:url] and options[:path].

While the ideal path for :timestamp is unclear to me as I'm not deeply familiar with the design philosophy behind Paperclip or the direction it wants to head in (aggressively fixing holes like this or supporting users who may rely on legacy behavior?), it seemed natural to me to expose the integer Attachment#updated_at value as an interpolation as an alternative to :timestamp. Epoch seconds don't carry any time zone information and are therefor immune to this issue, plus they're presumably faster to convert to than text, and fewer characters to hash to boot.

So my present implementation has a provisionally "improved" behavior for :timestamp (which should be backward compatible for everybody who didn't run into the bug personally), and a new :updated_at interpolation that merely exposes the existing Attachment instance method. I'd be eager for feedback on what if anything should be done with that stuff.

The caveats:

• I'm pretty new to testing, and this was definitely my first time writing tests using shoulda and test/unit. Feel free to slap me for transgressions in style and abuse of mocks/stubs/fakes/dummies.
• I tried to follow house style to the extent I could derive it, but I'm sure I got it wrong in places.
• I'm not sure what went on with the appraisal gemfile.locks reverting to an earlier version dependency on Rails -- perhaps this is expected? Again, never used appraisal before either. Tips would be more than welcome.

I'd be very excited if you see some or all of these changes as valuable to the Paperclip mainstream, and am happy to iterate this into something worthy of pulling.

Thanks a lot!

-john