An earlier version of post contained errors about the signing of binaries. Thanks to Guillaume Scerri for pointing them out.
On the first day of the summer school on secure and trustworthy computing, considerable time was dedicated to speakers from Intel talking about security-related CPU features. SGX is an upcoming such feature. Even though the introduction seemed to be more directed towards programmers rather than security researchers or even cryptographers, I was left with the following insights.
The core idea of SGX is to obscure the memory of an application from the operating system. Motivated by possible flaws in the OS, the applications themselves are separated from the OS. This is achieved by encrypting the application's memory using a key generated by the CPU, which never leaves the CPU. It is unclear to me whether SGX will hide memory access patterns of the application. However, this probably can be achieved using oblivious RAM within the application. Furthermore, I/O will be still be controlled by the OS, thus leaving the possibility of keylogging etc.
In order to convince a remote party that the binary is indeed running on SGX, processors will contain a private key to sign the binary being executed, with the corresponding public key being provided by the manufacturer. It is not clear whether the private key will be global or per CPU. A private key in the CPU would make it easy to also have the binaries encrypted, but it seems that this is not planned at the moment.
With SGX, Intel obviously targets cloud applications in an attempt to restrict the cloud provider's access to data in the cloud. Since there are no processors with SGX yet, it is hard to estimate the cost and the feasibility for SGX on mobile processors. However, if SGX eventually comes to mobile processors, rooting or jail-breaking becomes somewhat useless because the power of the OS will be restricted. This could be seen as another step of an evolution that went from proprietary software computing locally on personal data to personal data being held in the cloud. With SGX available on all processors, an app provider could exercise almost complete control over all cloud and mobile instances of the app, given the trustworthiness of the processor manufacturer. As a result, apps might effectively become black boxes with all interfaces defined by the app provider.
The second talk on SGX suggested its use as a replacement for multiparty computation as follows: The parties agree on a piece of software executing the computation. The software is then run using SGX and generates a public key, which the parties use to encrypt their inputs. After running the computation, the software only outputs the previously agreed results to the parties. One can argue that the security should be similar to the security of MPC because one has to trust the processor to execute the local computation correctly and without leakage to the adversary. Therefore, one might as well trust the processor with SGX to execute the software correctly while preserving confidentiality by encrypting the memory. However, this scheme introduces the CPU manufacturer as a third party possibly knowing the private key, which is used to confirm that the software runs on SGX. The manufacturer is inherently trusted not to collude with any party. More concisely, one might ask: You trust your CPU. Do you also trust the manufacturer?