- Chromium OS has been designed from the ground up with security in mind.
- Security is not a one-time effort, but rather an iterative process that must be focused on for the life of the operating system.
- The goal is that, should either the operating system or the user detect that the system has been compromised, an update can be initiated, and—after a reboot—the system will have been returned to a known good state.
- Chromium OS security strives to protect against an opportunistic adversary through a combination of system hardening, process isolation, continued web security improvements in Chromium, secure autoupdate, verified boot, encryption, and intuitive account management.
We have made a concerted effort to provide users of Chromium OS-based devices with a system that is both practically secure and easy to use. To do so, we've followed a set of four guiding principles:
- The perfect is the enemy of the good.
- Deploy defenses in depth.
- Make devices secure by default.
- Don't scapegoat our users.
In the rest of this document, we first explain these principles and discuss some expected use cases for Chromium OS devices. We then give a high-level overview of the threat model against which we will endeavor to protect our users, while still enabling them to make full use of their cloud device.
The perfect is the enemy of the good. No security solution is ever perfect. Mistakes will be made, there will be unforeseen interactions between multiple complex systems that create security holes, and there will be vulnerabilities that aren't caught by pre-release testing. Thus, we must not allow our search for some mythical perfect system to stop us from shipping something that is still very good. Deploy defenses in depth. In light of our first principle, we will deploy a variety of defenses to act as a series of stumbling blocks for the attacker. We will make it hard to get into the system, but assume that the attacker will. We'll put another layer of defenses in place to make it difficult to turn a user account compromise into root or a kernel exploit. Then, we'll also make it difficult for an attacker to persist their presence on the system by preventing them from adding an account, installing services, or re-compromising the system after reboot. Make it secure by default. Being safe is not an advanced or optional feature. Until now, the security community has had to deploy solutions that cope with arbitrary software running on users' machines; as a result, these solutions have often cost the user in terms of system performance or ease-of-use. Since we have the advantage of knowing which software should be running on the device at all times, we should be better able to deploy solutions that leave the user's machine humming along nicely. Don't scapegoat our users. In real life, people assess their risk all the time. The Web is really a huge set of intertwined, semi-compatible implementations of overlapping standards. Unsurprisingly, it is difficult to make accurate judgments about one's level of risk in the face of such complexity, and that is not our users' fault. We're working to figure out the right signals to send our users, so that we can keep them informed, ask fewer questions, require them to make decisions only about things they comprehend, and be sure that we fail-safe if they don't understand a choice and just want to click and make it go away.
We are initially targeting the following use cases with Chromium OS devices:
- Computing on the couch
- Use as a lightweight, secondary work computer
- Borrowing a device for use in coffee shops and libraries
- Sharing a second computer among family members
Targeting these goals dictates several security-facing requirements:
- The owner should be able to delegate login rights to users of their choice.
- The user can manage their risk with respect to data loss, even in the face of device loss or theft.
- A user's data can't be exposed due to the mistakes of other users on the system.
- The system provides a multi-tiered defense against malicious websites and other network-based attackers.
- Recovering from an attack that replaces or modifies system binaries should be as simple as rebooting.
- In the event of a security bug, once an update is pushed, the user can reboot and be safe.
When designing security technology for Chromium OS systems, we consider two different kinds of adversaries:
- An opportunistic adversary
- A dedicated adversary
The opportunistic adversary is just trying to compromise an individual user's machine and/or data. They are not targeting a specific user or enterprise, and they are not going to steal the user's machine to obtain the user's data. The opportunistic adversary will, however, deploy attacks designed to lure users on the web to sites that will compromise their machines or to web apps that will try to gain unwarranted privileges (webcam access, mic access, etc). If the opportunistic adversary does steal a device and the user's data is in the clear, the opportunistic adversary may take it.
The dedicated adversary may target a user or an enterprise specifically for attack. They are willing to steal devices to recover data or account credentials (not just to re-sell the device to make money). They are willing to deploy DNS or other network-level attacks to attempt to subvert the Chromium OS device login or update processes. They may also do anything that the opportunistic adversary can do.
For version 1.0, we are focusing on dangers posed by opportunistic adversaries. We further subdivide the possible threats into two different classes of attacks: remote system compromise and device theft.
There are several vectors through which an adversary might try to compromise a Chromium OS device remotely: an exploit that gives them control of one of the Chromium-based browser processes, an exploit in a plugin, tricking the user into giving a malicious web app unwarranted access to HTML5/Extension APIs, or trying to subvert our autoupdate process in order to get some malicious code onto the device.
As in any good security strategy, we wish to provide defense in depth: mechanisms that try to prevent these attacks and then several more layers of protection that try to limit how much damage the adversary can do provided that he's managed to execute one of these attacks. The architecture of Chromium browsers provides us with some very nice process isolation already, but there is likely more that we can do.
The lowest level of our security strategy involves a combination of OS-level protection mechanisms and exploit mitigation techniques. This combination limits our attack surface, reduces the the likelihood of successful attack, and reduces the usefulness of successful user-level exploits. These protections aid in defending against both opportunistic and dedicated adversaries. The approach designed relies on a number of independent techniques:
- Process sandboxing
- Mandatory access control implementation that limits resource, process, and kernel interactions
- Control group device filtering and resource abuse constraint
- Chrooting and process namespacing for reducing resource and cross-process attack surfaces
- Media device interposition to reduce direct kernel interface access from Chromium browser and plugin processes
- Toolchain hardening to limit exploit reliability and success
- NX, ASLR, stack cookies, etc
- Kernel hardening and configuration paring
- Additional file system restrictions
- Read-only root partition
- tmpfs-based /tmp
- User home directories that can't have executables, privileged executables, or device nodes
- Longer term, additional system enhancements will be pursued, like driver sandboxing
Detailed discussion can be found in the System Hardening design document.
The more modular the browser is, the easier it is for the Chromium OS to separate functionality and to sandbox different processes. Such increased modularity would also drive more efficient IPC within Chromium. We welcome input from the community here, both in terms of ideas and in code. Potential areas for future work include:
- All plugins should run as independent processes. We can then apply OS-level sandboxing techniques to limit their abilities. Approved plugins could even have Mandatory Access Control (MAC) policies generated and ready for use.
- Standalone network stack
- Chromium browsers already sandbox media and HTML parsers.
- If the HTTP and SSL stacks were isolated in independent processes, robustness issues with HTTP parsing or other behavior could be isolated. In particular, if all SSL traffic used one process while all plaintext traffic used another, we would have some protection from unauthenticated attacks leading to full information compromise. This can even be beneficial for cookie isolation, as the HTTP-only stack should never get access to cookies marked "secure." It would even be possible to run two SSL/HTTP network stacks—one for known domains based on a large whitelist and one for unknown domains. Alternately, it could be separated based on whether a certificate exception is required to finalize the connection.
- Per-domain local storage
- If it were possible to isolate renderer access per domain, then access to local storage services could similarly by isolated—at a process level. This would mean that a compromise of a renderer that escapes the sandbox would still not be guaranteed access to the other stored data unless the escape vector were a kernel-level exploit.
As we enable web applications to provide richer functionality for users, we are increasing the value of web-based exploits, whether the attacker tricks the browser into giving up extra access or the user into giving up extra access. We are working on multiple fronts to design a system that allows Chromium OS devices to manage access to new APIs in a unified manner, providing the user visibility into the behavior of web applications where appropriate and an intuitive way to manage permissions granted to different applications where necessary.
- User experience
- The user experience should be orthogonal to the policy enforcement mechanism inside the Chromium browser and, ideally, work the same for HTML5 APIs and Google Chrome Extensions.
- HTML5/Open Web Platform APIs and Google Chrome Extensions
- We're hopeful that we can unify the policy enforcement mechanisms/user preference storage mechanisms across all of the above.
- We are developing a multi-tiered sandboxing strategy, leveraging existing Chromium browser sandboxing technology and some of the work we discuss in our System Hardening design document.
- Long term, we will work with other browser vendors to make plugins more easily sandboxed.
- Full-screen mode in some plugins could allow an attacker to mock out the entire user experience of a Chromium OS device. We are investigating a variety of mitigation strategies in this space.
As HTML5 features like persistent workers move through the standards process, we must ensure that we watch for functionality creeping in that can poke holes in our security model and take care to handle it appropriately.
Attacks against the autoupdate process are likely to be executed by a dedicated adversary who would subvert networking infrastructure to inject a fake autoupdate with malicious code inside it. That said, a well supported opportunistic adversary could attempt to subvert the update process for many users simultaneously, so we should address this possibility here. (For more on this subject, also see the File System/Autoupdate design document.)
- Signed updates are downloaded over SSL.
- Version numbers of updates can't go backwards.
- The integrity of each update is verified on subsequent boot, using our Verified Boot process, described below.
Verified boot provides a means of getting cryptographic assurances that the Linux kernel, non-volatile system memory, and the partition table are untampered with when the system starts up. This approach is not "trusted boot" as it does not depend on a TPM device or other specialized processor features. Instead, a chain of trust is created using custom read-only firmware that performs integrity checking on a writable firmware. The verified code in the writable firmware then verifies the next component in the boot path, and so on. This approach allows for more flexibility than traditional trusted boot systems and avoids taking ownership away from the user. The design is broken down into two stages:
- Firmware-based verification
(for details, see the Firmware Boot and
- Read-only firmware checks writable firmware with a permanently stored key.
- Writable firmware then checks any other non-volatile memory as well as the bootloader and kernel.
- If verification fails, the user can either bypass checking or boot to a safe recovery mode.
- Kernel-based verification
(for details, see the Verified
Boot design document)
- This approach extends authenticity and integrity guarantees to files and metadata on the root file system.
- All access to the root file system device traverses a transparent layer which ensure data block integrity.
- Block integrity is determined using cryptographic hashes stored after the root file system on the system partition.
- All verification is done on-the-fly to avoid delaying system startup.
- The implementation is not tied to the firmware-based verification and may be compatible with any trusted kernel.
When combined, the two verification systems will perform as follows:
- Detects changes at boot-time
- Files, or read-write firmware, changed by an opportunistic attacker with a bootable USB drive will be caught on reboot.
- Changes performed by a successful runtime attack will also be detected on next reboot.
- Provides a secure recovery path so that new installs are safe from past attacks.
- Doesn't protect against
- Dedicated attackers replacing the firmware.
- Run-time attacks: Only code loaded from the file system is verified. Running code is not.
- Persistent attacks run by a compromised Chromium browser: It's not possible to verify the browser's configuration as safe using this technique.
It's important to note that at no point is the system restricted to code from the Chromium project; however, if Google Chrome OS is used, additional hardware-supported integrity guarantees can be made.
We do not intend to brick devices that we believe to be hacked. If we can reliably detect this state on the client, we should just initiate an update and reboot. We could try to leverage the abuse detection and mitigation mechanisms in the Google services that people are using from their Chromium OS devices, but it seems more scalable to allow each service to continue handling these problems on its own.
A stolen device is likely to have a higher value to a dedicated adversary than to an opportunistic adversary. An opportunistic adversary is more likely to reset the device for resale, or try to log in to use the device for himself.
The challenges here are myriad:
- We want to protect user data while also enabling users to opt-in to auto-login.
- We want to protect user data while also allowing users to share the device.
- We especially want to protect user credentials without giving up offline login, auto-login, and device sharing.
- Disk encryption can have real impact on battery life and performance speed.
- The attacker can remove the hard drive to circumvent OS-level protections.
- The attacker can boot the device from a USB device.
Users shouldn't need to worry about the privacy of their data if they forget their device in a coffee shop or share it with their family members. The easiest way to protect the data from opportunistic attackers is to ensure that it is unreadable except when it is in use by its owner.
The Protecting Cached User Data design document provides details on data protection. Key requirements for protecting cached user data (at rest) are as follows:
- Each user has their own encrypted store.
- All user data stored by the operating system, browser, and any plugins are encrypted.
- Users cannot access each other's data on a shared device.
- The system does not protect against attacks while a user is logged in.
- The system will attempt to protect against memory extraction (cold boot) attacks when additional hardware support arrives.
- The system does not protect against root file system tampering by a dedicated attacker (verified boot helps there).
Preventing the adversary from logging in to the system closes one easy pathway to getting the machine to execute code on their behalf. That said, many want this device to be just as sharable as a Google Doc. How can we balance these questions, as well as take into account certain practical concerns? These issues are discussed at length in the User Accounts and Management design document, with some highlights below.
- What are the practical concerns?
- Whose wifi settings do you use? Jane's settings on Bob's device in Bob's house don't work.
- But using Bob's settings no matter where the device is doesn't work either.
- And if Bob's device is set up to use his enterprise's wifi, then it's dangerous if someone steals it!
- "The owner"
- Each device has one and only one owner.
- User preferences are distinct from system settings.
- System settings, like wifi, follow the owner of a device.
- The owner can whitelist other users, who can then log in.
- The user experience for this feature should require only a few clicks or keystrokes.
- Promiscuous mode
- The owner can opt in to a mode where anyone with a Google account can log in.
- Guest mode
- Users can initiate a completely stateless session, which does not sync or cache data.
- All system settings would be kept out of this session, including networking config.
For design details, see the Login design document.
At a high level, here is how Chromium OS devices authenticate users:
- Try to reach Google Accounts online.
- If the service can't be reached, attempt to unwrap the locally-stored, TPM-wrapped keys we use for per-user data encryption. Successful unwrap means successful login.
- For every successful online login, use a salted hash of the password to wrap a fresh encryption key using the TPM.
There are a number of important convenience features around authentication that we must provide for users, and some consequences of integrating with Google Accounts we must deal with. These all have security consequences, and we discuss these (and potential mitigation) here. Additionally, we are currently working through the various tradeoffs of supporting non-Google OpenID providers as an authentication backend. CAPTCHAs
Rather than strictly rate limiting failed authentication attempts, Google Accounts APIs respond with CAPTCHAs if our servers believe an attack is underway. We do not want users to face CAPTCHAs to log in to their device; if the user has correctly provided their credentials, they should be successfully logged in. Furthermore, including HTML rendering code in our screen locker would introduce more potential for crashing bugs, which would give an attacker an opportunity to access the machine. That said, we cannot introduce a vector by which attackers can brute force Google Accounts. To work around this right now, we do offline credential checking when unlocking the screen and ignore problems at login time, though we realize this is not acceptable long-term and are considering a variety of ways to address this issue in time for V1.
As we discuss in the Login design document, we want to provide an SSO experience on the web for our users. Upon login, we decrypt the user's profile and perform a request for her Google Accounts cookies in the background. As a result, her profile gets fresh cookies and she is logged in to all her Google services.
When a user turns on auto-login, they are asserting that they wish this device to be trusted as though it had the user's credentials at all times; however, we don't want to store actual Google Account credentials on the device—doing so would expose them to offline attack unnecessarily. We also don't want to rely on an expiring cookie; auto-login would "only work sometimes," which is a poor user experience. We would like to store a revokable credential on the device, one that can be exchanged on-demand for cookies that will log the user in to all of their Google services. We're considering using an OAuth token for this purpose.
Biometrics, smartcards and Bluetooth
We expect to keep an eye on biometric authentication technologies as they continue to become cheaper and more reliable, but at this time we believe that the cost/reliability tradeoff is not where it needs to be for our target users. We expect these devices to be covered in our users' fingerprints, so a low-cost fingerprint scanner could actually increase the likelihood of compromise. We were able to break into one device that used facial recognition authentication software just by holding it up to the user's photo. Bluetooth adds a whole new software stack to our login/screenlocker code that could potentially be buggy, and the security of the pairing protocol has been criticized in the past. Smart cards and USB crypto tokens are an interesting technology, but we don't want our users to have to keep track of a physically distinct item just to use their device.
For third-party sites, we would like to provide credential generation and synced, cloud-based storage.
In this document, we have aimed only to summarize the wide-ranging efforts we are undertaking to secure Chromium OS at all levels. For more detail, please read the rest of the design documents.