When you write code to parse, evaluate, or otherwise handle untrustworthy inputs from the Internet — which is almost everything we do in a web browser! — we like to follow a simple rule to make sure it's safe enough to do so. The Rule Of 2 is: Pick no more than 2 of
(drawing source here)
When code that handles untrustworthy inputs at high privilege has bugs, the resulting vulnerabilities are typically of Critical or High severity. (See our Severity Guidelines.) We'd love to reduce the severity of such bugs by reducing the amount of damage they can do (lowering their privilege), avoiding the various types of memory corruption bugs (using a safe language), or reducing the likelihood that the input is malicious (asserting the trustworthiness of the source).
For the purposes of this document, our main concern is reducing (and hopefully, ultimately eliminating) bugs that arise due to memory unsafety. A recent study by Matt Miller from Microsoft Security states that “~70% of the vulnerabilities addressed through a security update each year continue to be memory safety issues”. A trip through Chromium's bug tracker will show many, many vulnerabilities whose root cause is memory unsafety. (As of March 2019, only about 5 of 130 public Critical-severity bugs are not obviously due to memory corruption.)
Security engineers in general, very much including Chrome Security Team, would like to advance the state of engineering to where memory safety issues are much more rare. Then, we could focus more attention on the application-semantic vulnerabilities. 😊 That would be a big improvement.
Some definitions are in order.
Untrustworthy inputs are inputs that
If there were an input type so simple that it were straightforward to write a memory-safe handler for it, we wouldn‘t need to worry much about where it came from for the purposes of memory safety, because we’d be sure we could handle it. We would still need to treat the input as untrustworthy after parsing, of course.
Unfortunately, it is very rare to find a grammar trivial enough that we can trust ourselves to parse it successfully or fail safely. (But see Normalization for a potential example.) Therefore, we do need to concern ourselves with the provenance of such inputs.
Any arbitrary peer on the Internet is an untrustworthy source, unless we get some evidence of its trustworthiness (which includes at least a strong assertion of the source's identity). When we can know with certainty that an input is coming from the same source as the application itself (e.g. Google in the case of Chrome, or Mozilla in the case of Firefox), and that the transport is integrity-protected (such as with HTTPS), then it can be acceptable to parse even complex inputs from that source. It's still ideal, where feasible, to reduce our degree of trust in the source — such as by parsing the input in a sandbox.
High privilege is a relative term. The very highest-privilege programs are the computer‘s firmware, the bootloader, the kernel, any hypervisor or virtual machine monitor, and so on. Below that are processes that run as an OS-level account representing a person; this includes the Chrome browser process. We consider such processes to have high privilege. (After all, they can do anything the person can do, with any and all of the person’s valuable data and accounts.)
Processes with slightly reduced privilege include (as of March 2019) the GPU process and (hopefully soon) the network process. These are still pretty high-privilege processes. We are always looking for ways to reduce their privilege without breaking them.
Turning a stream of bytes into a structured object is hard to do correctly and safely. For example, turning a stream of bytes into a sequence of Unicode code points, and from there into an HTML DOM tree with all its elements, attributes, and metadata, is very error-prone. The same is true of QUIC packets, video frames, and so on.
Whenever the code branches on the byte values it's processing, the risk increases that an attacker can influence control flow and exploit bugs in the implementation.
Although we are all human and mistakes are always possible, a function that does not branch on input values has a better chance of being free of vulnerabilities. (Consider an arithmetic function, such as SHA-256, for example.)
Chrome Security Team will generally not approve landing a CL or new feature that involves all 3 of untrustworthy inputs, unsafe language, and high privilege. To solve this problem, you need to get rid of at least 1 of those 3 things. Here are some ways to do that.
Also known as sandboxing, privilege reduction means running the code in a process that has had some or many of its privileges revoked.
When appropriate, try to handle the inputs in a renderer process that is Site Isolated to the same site as the inputs come from. Take care to validate the parsed (processed) inputs in the browser, since only the browser can trust itself to validate and act on the meaning of an object.
Equivalently, you can launch a sandboxed utility process to handle the data, and return a well-formed response back to the caller in an IPC message. See Safe Browsing's ZIP analyzer for an example. The Data Decoder Service facilitates this safe decoding process for several common data formats.
If you can be sure that the input comes from a trustworthy source, it can be OK to parse/evaluate it at high privilege in an unsafe language. A “trustworthy source” means that Chromium can cryptographically prove that the data comes from a business entity that you can or do trust (e.g. for Chrome, an Alphabet company).
Such cryptographic proof can potentially be obtained by:
Pinned TLS needs to meet all these criteria to be effective:
At present pinning is not enabled for all Chrome platforms. On other platforms, pinning may be disabled or rendered ineffective by enterprise security products. It is generally much better to use the Component Updater.
One common pattern is to deliver a cryptographic hash of some content via such a trustworthy channel, but deliver the content itself via an untrustworthy channel. So long as the hash is properly verified, that's fine.
You can ‘defang’ a potentially-malicious input by transforming it into a normal or minimal form, usually by first transforming it into a format with a simpler grammar. We say that all data, file, and wire formats are defined by a grammar, even if that grammar is implicit or only partially-specified (as is so often the case). A data format with a particularly simple grammar is
SkPixmap. (The ‘grammar’ is represented by the private data fields: a region of raw pixel data, the size of that region, and simple metadata (
SkImageInfo) about how to interpret the pixels.)
It's rare to find such a simple grammar for input formats, however.
For example, consider the PNG image format, which is complex and whose C implementation has suffered from memory corruption bugs in the past. An attacker could craft a malicious PNG to trigger such a bug. But if you transform the image into a format that doesn‘t have PNG’s complexity (in a low-privilege process, of course), the malicious nature of the PNG ‘should’ be eliminated and then safe for parsing at a higher privilege level. Even if the attacker manages to compromise the low-privilege process with a malicious PNG, the high-privilege process will only parse the compromised process' output with a simple, plausibly-safe parser. If that parse is successful, the higher-privilege process can then optionally further transform it into a normalized, minimal form (such as to save space). Otherwise, the parse can fail safely, without memory corruption.
The trick of this technique lies in finding a sufficiently-trivial grammar, and committing to its limitations.
Another good approach is to
That way, the higher-privileged process need only process objects adhering to a well-defined, generally low-complexity grammar. This is a big part of why we like for Mojo messages to use structured types.
For example, it should be safe enough to convert a PNG to an
SkBitmap in a sandboxed process, and then send the
SkBitmap to a higher-privileged process via IPC. Although there may be bugs in the IPC message deserialization code and/or in Skia's
SkBitmap handling code, we consider this safe enough for a few reasons:
Ultimately this process results in parsing significantly simpler grammars. (PNG → Mojo +
SkBitmap in this case.)
(We have to accept the risk of memory safety bugs in Mojo deserialization because C++‘s high performance is crucial in such a throughput- and latency-sensitive area. If we could change this code to be both in a safer language and still have such high performance, that’d be ideal. But that's unlikely to happen soon.)
While less preferable to Mojo, we also similarly trust Protobuf for deserializing messages at high privilege from potentially untrustworthy senders. For example, Protobufs are sometimes embedded in Mojo IPC messages. It is always preferable to use a Mojo message where possible, though sometimes external constraints require the use of Protobuf. Note that this only applies to Protobuf as a container format; the data contained within a Protobuf must be handled according to this rule as well.
For an example of image processing, we have the pure-Java class BaseGifImage. On Android, where we can use Java and also face a particularly high cost for creating new processes (necessary for sandboxing), using Java to decode tricky formats can be a great approach. We do a similar thing with the pure-Java JsonSanitizer, to ‘vet’ incoming JSON in a memory-safe way before passing the input to the C++ JSON implementation.
On Android, many system APIs that are exposed via Java are not actually implemented in a safe language, and are instead just facades around an unsafe implementation. A canonical example of this is the BitmapFactory class, which is a Java wrapper around C++ Skia. These APIs are therefore not considered memory-safe under the rule.
Regular expressions (re2) using trustworthy patterns can be used at high privilege to match on untrustworthy input strings. This does not automatically turn the matched text or captured groups into safe values.
As discussed above in Normalization, there are some types that are considered “safe,” even though they are deserialized from an untrustworthy source, at high privilege, and in an unsafe language. These types are fundamental for passing data between processes using IPC, tend to have simpler grammar or structure, and/or have been audited or fuzzed heavily.
There are also classes in //base that internally hold simple values that represent potentially complex data, such as:
The deserialization of these is safe, though it is important to remember that the value itself is still untrustworthy (e.g. a malicious path trying to escape its parent using
We still have code that violates this rule. For example, Chrome's Omnibox still parses JSON in the browser process. Additionally, the networking process on Windows is (at present) unsandboxed by default, though there is ongoing work to change that default.