Cross-Device Flows: Security Best Current Practice

2.4.1. Example A1: Authorize Access to a Video Streaming Service (User-Transferred Session Data Pattern)

An end-user sets up a new smart TV and wants to connect it to their favorite streaming service. The TV displays a QR code that the user scans with their mobile phone. The user is redirected to the streaming service provider's web page and asked to enter their credentials to authorize the smart TV to access the streaming service. The user enters their credentials and grants authorization, after which the streaming service is available on the smart TV.¶

2.4.2. Example A2: Authorize Access to Productivity Services (User-Transferred Session Data Pattern)

An employee wants to access their files on an interactive whiteboard in a conference room. The interactive whiteboard displays a URL and a code. The user enters the URL on their personal computer and is prompted for the code. Once they enter the code, the user is asked to authenticate and authorize the interactive whiteboard to access their files. The user enters their credentials and authorizes the transaction and the interactive whiteboard retrieves their files and allows the user to interact with the content.¶

2.4.3. Example A3: Authorize Use of a Bike Sharing Scheme (User-Transferred Session Data Pattern)

An end-user wants to rent a bicycle from a bike sharing scheme. The bicycles are locked in bicylce racks on sidewalks throughout a city. To unlock and use a bicycle, the user scans a QR code on the bicycle using their mobile phone. Scanning the QR code redirects the user to the bicycle sharing scheme's authorization page where the user authenticates and authorizes payment for renting the bicycle. Once authorized, the bicycle sharing service unlocks the bicycle, allowing the user to use it to cycle around the city.¶

2.4.4. Example A4: Authorize a Financial Transaction (Backchannel-Transferred Session Pattern)

An end-user makes an online purchase. Before completing the purchase, they get a notification on their mobile phone, asking them to authorize the transaction. The user opens their app and authenticates to the service before authorizing the transaction.¶

2.4.5. Example A5: Add a Device to a Network (User-Transferred Authorization Data Pattern)

An employee is issued with a personal computer that is already joined to a network. The employee wants to add their mobile phone to the network to allow it to access corporate data and services (e.g., files and e-mail). The employee is logged-in on the personal computer where they initiate the process of adding their mobile phone to the network. The personal computer displays a QR code which authorizes the user to join their mobile phone to the network. The employee scans the QR code with their mobile phone and the mobile phone is joined to the network. The employee can start accessing corporate data and services on their mobile device.¶

2.4.6. Example A6: Remote Onboarding (User-Transferred Session Data Pattern)

A new employee is directed to an onboarding portal to provide additional information to confirm their identity on their first day with their new employer. Before activating the employee's account, the onboarding portal requests that the employee present a government issued ID, proof of a background check and proof of their qualifications. The onboarding portal displays a QR code, which the user scans with their mobile phone. Scanning the QR code invokes the employee's digital wallet on their mobile phone, and the employee is asked to present digital versions of an identity document (e.g., a driving license), proof of a background check by an identity verifier, and proof of their qualifications. The employee authorizes the release of the credentials and after completing the onboarding process, their account is activated.¶

2.4.7. Example A7: Application Bootstrap (User-Transferred Authorization Data Pattern)

An employee is signed into an application on their personal computer and wants to bootstrap the mobile application on their mobile phone. The employee initiates the cross-device flow and is shown a QR code in their application. The employee launches the mobile application on their phone and scans the QR code which results in the user being signed into the application on the mobile phone.¶

2.4.8. Example A8: Access a Productivity Application (User-Transferred Authorization Data Pattern)

A user is accessing a Computer Aid Design (CAD) application. When accessing the application, authorization data in the form of a 6 digit authorization code is sent to the user's mobile phone. The user views the 6 digit authorization code on their phone and enters it in the CAD application, after which the CAD application displays the user's most recent designs.¶

3.4.1. Example B1: Illicit access to a video streaming service (User-Transferred Session Data Pattern)

An attacker obtains a smart TV and attempts to access an online streaming service. The smart TV obtains a QR code from the authorization server and displays it on screen. The attacker copies the QR code and embeds it in an e-mail that is sent to a large number of recipients. The e-mail contains a message stating that the streaming service wants to thank them for their loyal support and by scanning the QR code, they will be able to add a bonus device to their account for no charge. One of the recipients open the e-mail and scan the QR code to claim the loyalty reward. The user performs multi-factor authentication, and when asked if they want a new device to be added to their account, they authorize the action. The attacker's device is now authorized to access the content and obtains an access and refresh token. The access token allows the attacker to access content and the refresh token allows the attacker to obtain fresh tokens whenever the access token expires.¶

The attacker scales up the attack by emulating a new smart TV, obtaining multiple QR codes and widening the audience it sends the QR code to. Whenever a recipient scans the QR code and authorizes the addition of a new device, the attacker obtains an access and refresh token, which they sell for a profit.¶

3.4.2. Example B2: Illicit Access to Productivity Services (User-Transferred Session Data Pattern)

An attacker emulates an enterprise application (e.g., an interactive whiteboard) and initiates a cross-device flow by requesting a user code and URL from the authorization server. The attacker obtains a list of potential victims and sends an e-mail informing users that their files will be deleted within 24 hours if they don't follow the link, enter the user code and authenticate. The e-mail reminds them that this is the third time that they have been notified and their last opportunity to prevent deletion of their work files. One or more employees respond by following the URL, entering the code and performing multi-factor authentication. Throughout the authentication experience, the user is interacting with a trusted user experience, re-enforcing the legitimacy of the request. Once these employees authorized access, the attacker obtains access and refresh tokens from the authorization server and uses it to access the users' files, perform lateral attacks to obtain access to other information and continuously refresh the session by requesting new access tokens. These tokens may be exfiltrated and sold to third parties.¶

3.4.3. Example B3: Illicit Access to Physical Assets (User-Transferred Session Data Pattern)

An attacker copies a QR code from a bicycle locked in a bicycle rack in a city, prints it on a label and places the label on a bicycle at the other end of the bicycle rack. A customer approaches the bicycle that contains the replicated QR code and scans the code and authenticates before authorizing payment for renting the bicycle. The bicycle rack unlocks the bicycle containing the original QR code and the attacker removes the bicycle before cycling down the street while the customer is left frustrated that the bicycle they were trying to use is not being unlocked [NYC.Bike]. The customer proceeds to unlock another bicycle and lodges a complaint with the bicycle renting company.¶

3.4.4. Example B4: Illicit Transaction Authorization (Backchannel-Transferred Session Pattern)

An attacker obtains a list of user identifiers for a financial institution and triggers a transaction request for each of the users on the list. The financial institution's authorization server sends push notifications to each of the users, requesting authorization of a transaction. The vast majority of users ignore the request to authorize the transaction, but a small percentage grants authorization by approving the transaction.¶

3.4.5. Example B5: Illicit Network Join (User-Transferred Authorization Data Pattern)

An attacker creates a message to all employees of a company, claiming to be from a trusted technology provider investigating a suspected security breach. They ask employees to send them the QR code typically used to join a new device to the network, along with detailed steps on how to obtain the QR code. The employee, eager to assist, initiates the process to add a new mobile device to the network. They authenticate to the network and obtain a QR code. They send the QR code to the attacker. The attacker scans the QR code and adds their own device to the network. They use this device access as an entry point and perform lateral moves to obtain additional privileges and access to restricted resources.¶

3.4.6. Example B6: Illicit Onboarding (User-Transferred Session Data Pattern)

An attacker initiates an employee onboarding flow and obtains a QR code from the onboarding portal to invoke a digital wallet and present a verifiable credential attesting to a new employee's identity. The attacker obtains a list of potential new employees and sends an e-mail informing them that it is time to present proof of their background check or government issued ID. The new employee scans the QR code, invokes their digital wallet and presents their credentials. Once the credentials are presented, the employee's account is activated. The employee portal accessed by the attacker to obtain the QR code displays a message to the attacker with instructions on how to access their account.¶

3.4.7. Example B7: Illicit Application Bootstrap (User-Transferred Authorization Data Pattern)

An attacker creates a message to all employees of a company, claiming to be from the company's IT service provider. They claim that they are trying to resolve an application performance issue and ask employees to send them the QR code typically used to transfer a session. The employee, eager to assist, initiates the process to transfer a session. They authenticate and obtain a QR code and then send the QR code to the attacker. The attacker scans the QR code with their mobile phone and access the users data and resources.¶

3.4.8. Example B8: Account Takeover (User-Transferred Session Data Pattern)

An attacker wants to use some website which requires presentation of a verifiable credential for authentication. The attacker creates a phishing website which will in real time capture log-in QR Codes from the original website and present these to the user. The attacker tries to get the user to use the phishing website using an e-mail campaign etc. The user scans the QR code on the phishing website, invokes their digital wallet and presents their credentials. Once the credentials are presented, the original session from the attackers device is authorized with the user's credentials.¶

3.4.9. Out of Scope

In all of the attack scenarios listed above, a user is misled or exploited. For other attacks, where the user is willingly colluding with the attacker, the threat model, security implications and potential mitigations are very different. For example, a cooperating user can bypass software mitigations on their device, share access to hardware tokens with the attacker, and install additional devices to forward radio signals to circumvent proximity checks.¶

This document only considers scenarios where a user does not collude with an attacker.¶

5.1.1. Establish Proximity

The unauthenticated channel between the Initiating Device and Authorization Device allows attackers to obtain a QR code or user code in one location and display it in another location. Consequently, proximity-enforced cross-device flows are more resistant to Cross-Device Consent Phishing attacks than proximity-less cross-device flows. Establishing proximity between the location of the Initiating Device and the Authorization Device limits an attacker's ability to launch attacks by sending the user or QR codes to large numbers of users that are geographically distributed. There are a couple of ways to establish proximity:¶

• Physical connectivity: This is a good indicator of proximity, but requires specific ports, cables and hardware and may be challenging from a user experience perspective or may not be possible in certain settings (e.g., when USB ports are blocked or removed for security purposes). Physical connectivity may be better suited to dedicated hardware like FIDO devices that can be used with protocols that are resistant to the exploits described in this document.¶

• Wireless proximity: Near Field Communications (NFC), Bluetooth Low Energy (BLE), and Ultra Wideband (UWB) services can be used to prove proximity between the two devices. NFC technology is widely deployed in mobile phones as part of payment solutions, but NFC readers are less widely deployed. BLE presents another alternative for establishing proximity, but may present user experience challenges when setting up. UWB standards such as IEEE 802.15.4 and the IEEE 802.15.4z-2020 Amendment 1 enable secure ranging between devices and allow devices to establish proximity relative to each other [IEEE802154].¶

• Shared network: Device proximity can be inferred by verifying that both devices are on the same network. This check may be performed by the authorization server by comparing the network addresses of the device where the code is displayed (Initiating Device) with that of the Authorization Device. Alternatively the check can be performed on the device, provided that the network address is available. This could be achieved if the authorization server encodes the Initiating Device's network address in the QR code and uses a digital signature to prevent tampering with the code. This does require the wallet to be aware of the countermeasure and effectively enforce it.¶

• Geo-location: Proximity can be established by comparing geo-location information derived from global navigation satellite-system (GNSS) co-ordinates or geolocation lookup of IP addresses and comparing proximity. Due to inaccuracies, this may require restrictions to be at a more granular level (e.g., same city, country, region or continent). Similar to the shared network checks, these checks may be performed by the authorization server or on the users device, provided that the information encoded in a QR code is integrity protected using a digital signature.¶

Depending on the risk profile and the threat model in which a system is operating, it may be necessary to use more than one mechanism to establish proximity to raise the bar for any potential attackers.¶

Note: There are scenarios that require that an authorization takes place in a different location than the one in which the transaction is authorized. For example, there may be a primary and secondary credit card holder and both can initiate transactions, but only the primary holder can authorize it. There is no guarantee that the primary and secondary holders are in the same location at the time of the authorization. In such cases, proximity (or lack of proximity) may be an indicator of risk and the system may deploy additional controls (e.g., transaction value limits, transaction velocity limits) or use the proximity information as input to a risk management system.¶

Limitations: Proximity mechanisms make it harder to perform Cross-Device Consent Phishing (CDCP) attacks. However, depending on how the proximity check is performed, an attacker may be able to circumvent the protection: The attacker can use a VPN to simulate a shared network or spoof a GNSS position. For example, the attacker can try to request the location of the end-user's Authorization Device through browser APIs and then simulate the same location on their Initiating Device using standard debugging features available on many platforms.¶

5.1.2. Short Lived/Timebound QR or User Codes

The impact of an attack can be reduced by making QR or user codes short lived. If an attacker obtains a short-lived code, the duration during which the unauthenticated channel can be exploited is reduced, potentially increasing the cost of a successful attack.¶

Limitations: There is a practical limit to how short a user code can be valid due to network latency and user experience limitations (time taken to enter a code, or incorrectly entering a code). More sophisticated Cross-Device Consent Phishing attacks counter the effectiveness of short lived codes by convincing a user to respond to a phishing e-mail and only request the QR or user code once the user clicks on the link in the phishing e-mail [Exploit6].¶

5.1.3. One-Time or Limited Use Codes

By enforcing one-time use or limited use of user or QR codes, the authorization server can limit the impact of attacks where the same user code or QR code is sent to multiple victims. One-time use may be achieved by including a nonce or date-stamp in the user code or QR code which is validated by the authorization server when the user scans the QR code against a list of previously issued codes.¶

Limitations: Enforcing one-time use may be difficult in large globally distributed systems with low latency requirements, in which case short lived tokens may be more practical. One-time use codes may also have an impact on the user experience. For example, a user may enter a code, but their session may be interrupted before the access request is completed. If the code is a one-time use code, they would need to restart the session and obtain a new code since they won't be allowed to enter the same code a second time. As a result, it may be practical to allow the same code to be presented a small number of times.¶

5.1.4. Unique Codes

By issuing unique user or QR codes, an authorization server can detect if the same codes are being repeatedly submitted. This may be interpreted as anomalous behavior and the authorization server may choose to decline issuing access and refresh tokens if it detects the same codes being presented repeatedly. This may be achieved by maintaining a deny list that contains QR codes or user codes that were previously used. The authorization server may use a sliding window equal to the lifetime of a token if short lived/timebound tokens are used (see Short Lived/Timebound Codes). This will limit the size of the deny list.¶

Limitations: Maintaining a deny list of previously redeemed codes, even for a sliding window, may have an impact on the latency of globally distributed systems. One alternative is to segment user codes by geography or region and maintain local deny lists.¶

5.1.5. Content Filtering

Attackers exploit the unauthenticated channel by changing the context of the user code or QR code and then sending a message to a user (e-mail, text, instant messaging etc). By deploying content filtering (e.g., anti-spam filter), these messages can be blocked and prevented from reaching the end-users. It may be possible to fine-tune content filtering solutions to detect artefacts like QR codes or user codes that are included in a message that is sent to multiple recipients in the expectation that at least one of the recipients will be convinced by the message and grant authorization to access restricted resources.¶

Limitations: Some scenarios may require legitimate re-transmission of user, QR and authorization data (e.g. retries). To prevent the disruption of legitimate scenarios, content filters may use a threshold and allow a limited number of messages with the same QR or user codes to be transmitted before interrupting the delivery of those messages. Content filtering may also be fragmented across multiple communications systems and channels (e-mail, messaging, text etc), making it harder to detect or interrupt attacks that are executed over multiple channels, unless here is a high degree of integration between content filtering systems.¶

5.1.6. Detect and Remediate

The authorization server may be able to detect misuse of the codes due to repeated use as described in Unique Codes, as an input from a content filtering engine as described in Content Filtering, or through other mechanisms such as reports from end users. If an authorization server determines that a user code or QR code is being used in an attack it may choose to invalidate all tokens issued in response to these codes and make that information available through a token introspection endpoint (see [RFC7662]. In addition it may notify resource servers to stop accepting these tokens or to terminate existing sessions associated with these tokens using Continuous Access Evaluation Protocol (CAEP) messages [CAEP] using the Shared Signals Framework (SSF) [SSF] framework or an equivalent notification system.¶

Limitations: Detection and remediation requires that resource servers are integrated with security eventing systems or token introspection services. This may not always be practical for existing systems and may need to be targeted to the most critical resource services in an environment.¶

5.1.7. Trusted Devices

If an attacker is unable to initiate the protocol, they are unable to obtain a QR code or user code that can be leveraged for the attacks described in this document. By restricting the protocol to only be executed on devices trusted by the authorization server, it prevents attackers from using arbitrary devices, or by mimicking devices to initiate the protocol. Trusted devices include devices that are pre-registered with the authorization server or are subject to device management policies. Device management policies may enforce patching, version updates, on-device anti-malware deployment, revocation status and device location amongst others. Trusted devices may have their identities rooted in hardware (e.g., a TPM or equivalent technology). By only allowing trusted devices to initiate cross-device flows, it requires the attacker to have access to such a device and maintain access in a way that does not result in the device's trust status from being revoked.¶

Limitations: An attacker may still be able to obtain access to a trusted device and use it to initiate authorization requests, making it necessary to apply additional controls and integrating with other threat detection and management systems that can detect suspicious behaviour such as repeated requests to initiate authorization or high volume of service activation on the same device.¶

5.1.8. Trusted Networks

An attacker can be prevented from initiating a cross-device flow protocol by only allowing the protocol to be initiated on a trusted network or within a security perimeter (e.g., a corporate network). A trusted network may be defined as a set of IP addresses and joining the network is subject to security controls managed by the network operator, which may include only allowing trusted devices on the network, device management, user authentication and physical access policies and systems. By limiting protocol initiation to a specific network, the attacker needs to have access to a device on the network.¶

Limitations: Network level controls may not always be feasible, especially when dealing with consumer scenarios where the network may not be under control of the service provider. Even if it is possible to deploy network level controls, it should be used in conjunction with other controls outlined in this document to achieve defence in-depth.¶

5.1.9. Limited Scopes

Authorization servers may choose to limit the scopes they include in access tokens issued through cross-device flows where the unauthenticated channel between two devices are susceptible to being exploited. Including limited scopes lessens the impact in case of a successful attack. The decision about which scopes are included may be further refined based on whether the protocol is initiated on a trusted device or the user's location relative to the Initiating Device.¶

Limitations: Limiting scopes reduces the impact of a compromise, but does not avoid it. It should be used in conjunction with other mitigations described in this document.¶

5.1.10. Short Lived Tokens

Another mitigation strategy includes limiting the life of the access and refresh tokens. The lifetime can be lengthened or shortened, depending on the user's location, the resources they are trying to access or whether they are using a trusted device. Short lived tokens do not prevent or disrupt the attack, but serve as a remedial mechanism in case the attack succeeded.¶

Limitations: Short lived tokens reduces the time window during which an attacker can benefit from a successful attack. This is most effective for access tokens. However, once an attacker obtains a refresh token, they can continue to request new access tokens, as well as refresh tokens. Forcing the expiry of refresh tokens may cause the user to re-authorize an action more frequently, which results in a negative user experience.¶

5.1.11. Rate Limits

An attacker that engages in a scaled attack may need to request a large number of user codes (see exploit Example B1) or initiate a large number of authorization requests (see exploit Example B4) in a short period of time. An authorization server may apply rate limits to minimize the number of requests it would accept from a client in a limited time period.¶

Limitations: Rate limits are effective at slowing an attacker down and help to degrade scaled attacks, but do not prevent more targeted attacks that are executed with lower volumes and velocity. Therefore, it should be used along with other techniques to provide a defence-in-depth defence against cross-device attacks.¶

5.1.12. Sender-Constrained Tokens

Sender-constrained tokens limit the impact of a successful attack by preventing the tokens from being moved from the device on which the attack was successfully executed. This makes attacks where an attacker gathers a large number of access and refresh tokens on a single device and then sells them for profit more difficult, since the attacker would also have to export the cryptographic keys used to sender-constrain the tokens or be able to access them and generate signatures for future use. If the attack is being executed on a trusted device to a device with anti-malware, any attempts to exfiltrate tokens or keys may be detected and the device's trust status may be changed. Using hardware keys for sender-constraining tokens will further reduce the ability of the attacker to move tokens to another device.¶

Limitations: Sender-constrained tokens, especially sender-constrained tokens that require proof-of-posession, raise the bar for executing the attack and profiting from exfiltrating tokens. Although a software proof-of-posession key is better than no proof-of-posession key, an attacker may still exfiltrate the software key. Hardware keys are harder to exfiltrate, but come with additional implementation complexity. An attacker that controls the Initiating Device may still be able to excercise the key, even if it is in hardware. Consequently the main protection derived from sender-constrained tokens is preventing tokens from being moved from the Initiating Device to another device, thereby making it harder sell stolen tokens and profit from the attack.¶

5.1.13. User Experience

The user experience should preserve the context within which the protocols were initiated and communicate this clearly to the user when they are asked to authorize, authenticate or present a credential. In preserving the context, it should be clear to the user who invoked the flow, why it was invoked and what the consequence of completing the authorization, authentication or credential presentation. The user experience should reinforce the message that unless the user initiated the authorization request, or was expecting it, they should decline the request.¶

It should be clear to the user how to decline the request. To avoid accidental authorization grants, the "decline" option may be the default option or given similar prominence in the user experience as the "grant" option.¶

This information may be communicated graphically or in a simple message (e.g., "It looks like you are trying to access your files on a digital whiteboard in your city center office. Click here to grant access to your files. If you are not trying to access your files, you should decline this request and notify the security department").¶

The service may provide out-of-band reinforcement to the user on the context and conditions under which an authorization grant may be requested. For example if the service provider does not send e-mails with QR codes requesting users to grant authorization, this may be reinforced in marketing messages, in-app experiences and through anti-fraud awareness campaigns.¶

Limitations: Improvements to user experience on their own is unlikely to be sufficient and should be used in conjunction with other controls described in this document.¶

5.1.14. Authenticated flow

By requiring a user to authenticate on the Initiating Device with a phishing resistant authentication method before initiating a cross-device flow, the server can prevent an attacker from initiating a cross-device flow and obtaining QR codes or user codes. This prevents the attacker from obtaining a QR code or user code that they can use to mislead an unsuspecting user. This requires that the Initiating Device has sufficient input capabilities to support a phishing resistant authentication mechanism, which may in itself negate the need for a cross-device flow.¶

Limitations: Starting with an authenticated flow does not prevent the attacks described in Example B5: Illicit Network Join and Example B7: Illicit Application Bootstrap and it is recommended that additional mitigations described in this document is used if the cross-device flows are used in scenarios such as Example A5: Add a device to a network and Example A7: Application Bootstrap.¶

5.1.15. Practical Mitigation Summary

The practical mitigations described in this section can prevent the attacks from being initiated, disrupt attacks once they start or reduce the impact or remediate an attack if it succeeds. When combining one or more of these mitigations the overall security profile of a cross-device flow improves significantly. The following table provides a summary view of these mitigations:¶

5.2.1. IETF OAuth 2.0 Device Authorization Grant [RFC8628]:

5.2.2. OpenID Foundation Client Initiated Back-Channel Authentication (CIBA):

5.2.3. FIDO2/WebAuthn

5.2.4. Protocol Selection Summary

The FIDO Cross-Device Authentication (CDA) flow provides the best protection against attacks on the unauthenticated channel for cross device flows. It can be combined with OAuth 2.0 and OpenID Connect protocols for standards-based authorization and authentication flows. If FIDO2/WebAuthn support is not available, Client Initiated Backchannel Authentication (CIBA) provides an alternative, provided that there is a channel through which the authorization server can contact the end user. Examples of such a channel include device push notifications, e-mail or text messages which the user can access from their device. If CIBA is used, additional mitigations to enforce proximity and initiate transactions from trusted devices or trusted networks should be considered. The OAuth 2.0 Device Authorization Grant provides the most flexibility and has the lowest requirements on devices used, but it is recommended that it is only used when additional mitigations are deployed to prevent attacks that exploit the unauthenticated channel between devices.¶