Rell researchers at the Graz University of Technology found a way to read arbitrary memory over a network called NetSpectre. NetSpectre attacks have been shown to work over LAN and Google Cloud. It's slow, though likely too slow as NetSpectre is only able to leak at an incredibly low rate of 15~60 bits per hour depending upon the processor.
So yeah the speed of the attack is limited to 60 bits per hour. Intel was notified of the exploit on March 20th, 2018 and agreed to the disclosure date in July 2018. NetSpectre is a new network-based speculative attack vulnerability that doesn't require exploited code to be running on the target machine. There Register on the topic:
Establishing a network connection to a service running exploitable snippets of code should, in theory, be enough to very slowly discern the contents of application memory remotely. This requires precise timing and constant measurement, so noisy network environments, such as the internet, will hamper exploitation to some extent.
That's the first stage. The next step is to pull out interesting data rather than grab temporary variables and other inconsequential stuff lying around in a program's memory – a step that is non-trivial.
"We show that Spectre attacks do not require local code execution but can also be mounted remotely," said Michael Schwartz, one of the NetSpectre researchers, in an email to The Register. "Moreover, with the new covert channel, we show that Spectre does not necessarily require the cache to leak values."
The major catch, described in a paper titled NetSpectre: Read Arbitrary Memory over Network, is that this side-channel attack only leaks 15 bits per hour, or 60 bits an hour via an AVX-based covert channel, which means it could take days to find and gather privileged information such as an encryption key or authentication token.
Schwartz reckons this data leakage is something people should worry about, although, admitted that the speed at which it can be conducted is a limiting factor.
"Luckily, the speed is quite limited, which makes this attack mainly interesting for targeted attacks on high-value targets," he said. "If the system is fully patched against Spectre, including the new gadget variants we show in the paper, the attack should be prevented. However, we are still at the beginning of understanding how Spectre gadgets can look like, so this is not a problem that is trivial to solve."
Spectre attacks manipulate the branch prediction mechanisms used in modern CPUs' speculative execution engines to force the target process to access memory in a way that leaks privileged information. Today's processors rely on speculative execution to run software at high speed, predicting where the flow of the program will go ahead and priming themselves with code and data in anticipation. It is possible to discern the contents of memory that is otherwise out of sight by manipulating and observing the effects of this predictive execution.
For a remote Spectre attack, the targeted device must include code that performs an operation such as an reading through an array in a loop with a bounds check on each iteration. The exploit abuses design decisions within the processor microarchitecture to induce speculative execution, and discern the content of memory as a result. The paper, written by Michael Schwarz, Daniel Gruss, Martin Schwarzl, Moritz Lipp, and Stefan Mangard of Austria's Graz University of Technology, calls these code fragments "Spectre gadgets."
Red Hat says it has been working with the researchers and plans to publish details about the impact on its products, if any, in a blog post on Friday. "We have not identified any viable userspace Spectre gadget attacks but are actively auditing all of the daemons that listen over the network and the rest of the stack," said Jon Masters, chief Arm architect and computer microarchitecture lead at Red Hat, via Twitter.
So far, as with the other Spectre and Meltdown variants and sub-variants, no malware is exploiting these flaws in the wild, that we know of.