Security Considerations for WebRTC

Security for WebRTC Applications

Access to Local Devices As discussed in , allowing arbitrary sites to initiate calls violates the core Web security guarantee; without some access restrictions on local devices, any malicious site could simply bug a user. At minimum, then, it MUST NOT be possible for arbitrary sites to initiate calls to arbitrary locations without user consent. This immediately raises the question, however, of what should be the scope of user consent. In order for the user to make an intelligent decision about whether to allow a call (and hence their camera and microphone input to be routed somewhere), they must understand either who is requesting access, where the media is going, or both. As detailed below, there are two basic conceptual models:

You are sending your media to entity A because you want to talk to entity A (e.g., your mother).
Entity A (e.g., a calling service) asks to access the user's devices with the assurance that it will transfer the media to entity B (e.g., your mother).

In either case, identity is at the heart of any consent decision. Moreover, the identity of the party the browser is connecting to is all that the browser can meaningfully enforce; if you are calling A, A can simply forward the media to C. Similarly, if you authorize A to place a call to B, A can call C instead. In either case, all the browser is able to do is verify and check authorization for whoever is controlling where the media goes. The target of the media can of course advertise a security/privacy policy, but this is not something that the browser can enforce. Even so, there are a variety of different consent scenarios that motivate different technical consent mechanisms. We discuss these mechanisms in the sections below. It's important to understand that consent to access local devices is largely orthogonal to consent to transmit various kinds of data over the network (see ). Consent for device access is largely a matter of protecting the user's privacy from malicious sites. By contrast, consent to send network traffic is about preventing the user's browser from being used to attack its local network. Thus, we need to ensure communications consent even if the site is not able to access the camera and microphone at all (hence WebSockets's consent mechanism) and similarly, we need to be concerned with the site accessing the user's camera and microphone even if the data is to be sent back to the site via conventional HTTP-based network mechanisms such as HTTP POST.

Threats from Screen Sharing In addition to camera and microphone access, there has been demand for screen and/or application sharing functionality. Unfortunately, the security implications of this functionality are much harder for users to intuitively analyze than for camera and microphone access. (See for a full analysis.) The most obvious threats are simply those of "oversharing". I.e., the user may believe they are sharing a window when in fact they are sharing an application, or may forget they are sharing their whole screen, icons, notifications, and all. This is already an issue with existing screen sharing technologies and is made somewhat worse if a partially trusted site is responsible for asking for the resource to be shared rather than having the user propose it. A less obvious threat involves the impact of screen sharing on the Web security model. A key part of the Same-Origin Policy is that HTML or JS from site A can reference content from site B and cause the browser to load it, but (unless explicitly permitted) cannot see the result. However, if a Web application from a site is screen sharing the browser, then this violates that invariant, with serious security consequences. For example, an attacker site might request screen sharing and then briefly open up a new window to the user's bank or webmail account, using screen sharing to read the resulting displayed content. A more sophisticated attack would be to open up a source view window to a site and use the screen sharing result to view anti-cross-site request forgery tokens. These threats suggest that screen/application sharing might need a higher level of user consent than access to the camera or microphone.

Calling Scenarios and User Expectations While a large number of possible calling scenarios are possible, the scenarios discussed in this section illustrate many of the difficulties of identifying the relevant scope of consent.

Dedicated Calling Services The first scenario we consider is a dedicated calling service. In this case, the user has a relationship with a calling site and repeatedly makes calls on it. It is likely that rather than having to give permission for each call, the user will want to give the calling service long-term access to the camera and microphone. This is a natural fit for a long-term consent mechanism (e.g., installing an app store "application" to indicate permission for the calling service). A variant of the dedicated calling service is a gaming site (e.g., a poker site) which hosts a dedicated calling service to allow players to call each other. With any kind of service where the user may use the same service to talk to many different people, there is a question about whether the user can know who they are talking to. If I grant permission to calling service A to make calls on my behalf, then I am implicitly granting it permission to bug my computer whenever it wants. This suggests another consent model in which a site is authorized to make calls but only to certain target entities (identified via media-plane cryptographic mechanisms as described in and especially ). Note that the question of consent here is related to but distinct from the question of peer identity: I might be willing to allow a calling site to in general initiate calls on my behalf but still have some calls via that site where I can be sure that the site is not listening in.

Calling the Site You're On Another simple scenario is calling the site you're actually visiting. The paradigmatic case here is the "click here to talk to a representative" windows that appear on many shopping sites. In this case, the user's expectation is that they are calling the site they're actually visiting. However, it is unlikely that they want to provide a general consent to such a site; just because I want some information on a car doesn't mean that I want the car manufacturer to be able to activate my microphone whenever they please. Thus, this suggests the need for a second consent mechanism where I only grant consent for the duration of a given call. As described in , great care must be taken in the design of this interface to avoid the users just clicking through. Note also that the user interface chrome, which is the representation through which the user interacts with the user agent itself, must clearly display elements showing that the call is continuing in order to avoid attacks where the calling site just leaves it up indefinitely but shows a Web UI that implies otherwise.

Origin-Based Security Now that we have described the calling scenarios, we can start to reason about the security requirements. As discussed in , the basic unit of Web sandboxing is the origin, and so it is natural to scope consent to the origin. Specifically, a script from origin A MUST only be allowed to initiate communications (and hence to access the camera and microphone) if the user has specifically authorized access for that origin. It is of course technically possible to have coarser-scoped permissions, but because the Web model is scoped to the origin, this creates a difficult mismatch. Arguably, the origin is not fine-grained enough. Consider the situation where Alice visits a site and authorizes it to make a single call. If consent is expressed solely in terms of the origin, then on any future visit to that site (including one induced via a mash-up or ad network), the site can bug Alice's computer, use the computer to place bogus calls, etc. While in principle Alice could grant and then revoke the privilege, in practice privileges accumulate; if we are concerned about this attack, something else is needed. There are a number of potential countermeasures to this sort of issue.

Individual Consent: Ask the user for permission for each call.
Callee-oriented Consent: Only allow calls to a given user.
Cryptographic Consent: Only allow calls to a given set of peer keying material or to a cryptographically established identity.

Unfortunately, none of these approaches is satisfactory for all cases. As discussed above, individual consent puts the user's approval in the UI flow for every call. Not only does this quickly become annoying but it can train the user to simply click "OK", at which point the consent becomes useless. Thus, while it may be necessary to have individual consent in some cases, this is not a suitable solution for (for instance) the calling service case. Where necessary, in-flow user interfaces must be carefully designed to avoid the risk of the user blindly clicking through. The other two options are designed to restrict calls to a given target. Callee-oriented consent provided by the calling site would not work well because a malicious site can claim that the user is calling any user of their choice. One fix for this is to tie calls to a cryptographically established identity. While not suitable for all cases, this approach may be useful for some. If we consider the case of advertising, it's not particularly convenient to require the advertiser to instantiate an IFRAME on the hosting site just to get permission; a more convenient approach is to cryptographically tie the advertiser's certificate to the communication directly. We're still tying permissions to the origin here, but to the media origin (and/or destination) rather than to the Web origin. describes mechanisms which facilitate this sort of consent. Another case where media-level cryptographic identity makes sense is when a user really does not trust the calling site. For instance, I might be worried that the calling service will attempt to bug my computer, but I also want to be able to conveniently call my friends. If consent is tied to particular communications endpoints, then my risk is limited. Naturally, it is somewhat challenging to design UI primitives which express this sort of policy. The problem becomes even more challenging in multi-user calling cases.

Security Properties of the Calling Page Origin-based security is intended to secure against Web attackers. However, we must also consider the case of network attackers. Consider the case where I have granted permission to a calling service by an origin that has the HTTP scheme, e.g., <http://calling-service.example.com>. If I ever use my computer on an unsecured network (e.g., a hotspot or if my own home wireless network is insecure), and browse any HTTP site, then an attacker can bug my computer. The attack proceeds like this:

I connect to <http://anything.example.org/>. Note that this site is unaffiliated with the calling service.
The attacker modifies my HTTP connection to inject an IFRAME (or a redirect) to <http://calling-service.example.com>.
The attacker forges the response from <http://calling-service.example.com/> to inject JS to initiate a call to themselves.

Note that this attack does not depend on the media being insecure. Because the call is to the attacker, it is also encrypted to them. Moreover, it need not be executed immediately; the attacker can "infect" the origin semi-permanently (e.g., with a Web worker or a popped-up window that is hidden under the main window) and thus be able to bug me long after I have left the infected network. This risk is created by allowing calls at all from a page fetched over HTTP. Even if calls are only possible from HTTPS sites, if those sites include active content (e.g., JavaScript) from an untrusted site, that JavaScript is executed in the security context of the page . This could lead to compromise of a call even if the parent page is safe. Note: This issue is not restricted to pages which contain untrusted content. If any page from a given origin ever loads JavaScript from an attacker, then it is possible for that attacker to infect the browser's notion of that origin semi-permanently.

Communications Consent Verification As discussed in , allowing Web applications unrestricted network access via the browser introduces the risk of using the browser as an attack platform against machines which would not otherwise be accessible to the malicious site, for instance, because they are topologically restricted (e.g., behind a firewall or NAT). In order to prevent this form of attack as well as cross-protocol attacks, it is important to require that the target of traffic explicitly consent to receiving the traffic in question. Until that consent has been verified for a given endpoint, traffic other than the consent handshake MUST NOT be sent to that endpoint. Note that consent verification is not sufficient to prevent overuse of network resources. Because WebRTC allows for a Web site to create data flows between two browser instances without user consent, it is possible for a malicious site to chew up a significant amount of a user's bandwidth without incurring significant costs to themselves by setting up such a channel to another user. However, as a practical matter there are a large number of Web sites which can act as data sources, so an attacker can at least use downlink bandwidth with existing Web APIs. However, this potential DoS vector reinforces the need for adequate congestion control for WebRTC protocols to ensure that they play fair with other demands on the user's bandwidth.

ICE Verifying receiver consent requires some sort of explicit handshake, but conveniently we already need one in order to do NAT hole-punching. Interactive Connectivity Establishment (ICE) includes a handshake designed to verify that the receiving element wishes to receive traffic from the sender. It is important to remember here that the site initiating ICE is presumed malicious; in order for the handshake to be secure, the receiving element MUST demonstrate receipt/knowledge of some value not available to the site (thus preventing the site from forging responses). In order to achieve this objective with ICE, the Session Traversal Utilities for NAT (STUN) transaction IDs must be generated by the browser and MUST NOT be made available to the initiating script, even via a diagnostic interface. Verifying receiver consent also requires verifying the receiver wants to receive traffic from a particular sender, and at this time; for example, a malicious site may simply attempt ICE to known servers that are using ICE for other sessions. ICE provides this verification as well, by using the STUN credentials as a form of per-session shared secret. Those credentials are known to the Web application, but would need to also be known and used by the STUN-receiving element to be useful. There also needs to be some mechanism for the browser to verify that the target of the traffic continues to wish to receive it. Because ICE keepalives are indications, they will not work here. describes the mechanism for providing consent freshness.

Masking Once consent is verified, there still is some concern about misinterpretation attacks as described by Huang et al. . This does not seem like it is of serious concern with DTLS because the ICE handshake enforces receiver consent and there is little evidence of passive DTLS proxies of the type studied by Huang. However, because RTCWEB can run over TCP there is some concern that attackers might control the ciphertext by controlling the plaintext input to SCTP. This risk is only partially mitigated by the fact that the SCTP stack controls the framing of the packets. Note that in principle an attacker could exert some control over Secure Real-time Transport Protocol (SRTP) packets by using a combination of the WebAudio API and extremely tight timing control. The primary risk here seems to be carriage of SRTP over Traversal Using Relays around NAT (TURN) TCP. However, as SRTP packets have an extremely characteristic packet header it seems unlikely that any but the most aggressive intermediaries would be confused into thinking that another application-layer protocol was in use.

Backward Compatibility A requirement to use ICE limits compatibility with legacy non-ICE clients. It seems unsafe to completely remove the requirement for some check. All proposed checks have the common feature that the browser sends some message to the candidate traffic recipient and refuses to send other traffic until that message has been replied to. The message/reply pair must be generated in such a way that an attacker who controls the Web application cannot forge them, generally by having the message contain some secret value that must be incorporated (e.g., echoed, hashed into, etc.). Non-ICE candidates for this role (in cases where the legacy endpoint has a public address) include:

STUN checks without using ICE (i.e., the non-RTC-web endpoint sets up a STUN responder).
Use of the RTP Control Protocol (RTCP) as an implicit reachability check.

In the RTCP approach, the WebRTC endpoint is allowed to send a limited number of RTP packets prior to receiving consent. This allows a short window of attack. In addition, some legacy endpoints do not support RTCP, so this is a much more expensive solution for such endpoints, for which it would likely be easier to implement ICE. For these two reasons, an RTCP-based approach does not seem to address the security issue satisfactorily. In the STUN approach, the WebRTC endpoint is able to verify that the recipient is running some kind of STUN endpoint but unless the STUN responder is integrated with the ICE username/password establishment system, the WebRTC endpoint cannot verify that the recipient consents to this particular call. This may be an issue if existing STUN servers are operated at addresses that are not able to handle bandwidth-based attacks. Thus, this approach does not seem satisfactory either. If the systems are tightly integrated (i.e., the STUN endpoint responds with responses authenticated with ICE credentials), then this issue does not exist. However, such a design is very close to an ICE-Lite implementation (indeed, arguably is one). An intermediate approach would be to have a STUN extension that indicated that one was responding to WebRTC checks but not computing integrity checks based on the ICE credentials. This would allow the use of standalone STUN servers without the risk of confusing them with legacy STUN servers. If a non-ICE legacy solution is needed, then this is probably the best choice. Once initial consent is verified, we also need to verify continuing consent, in order to avoid attacks where two people briefly share an IP (e.g., behind a NAT in an Internet cafe) and the attacker arranges for a large, unstoppable, traffic flow to the network and then leaves. The appropriate technologies here are fairly similar to those for initial consent, though are perhaps weaker since the threats are less severe.

IP Location Privacy Note that as soon as the callee sends their ICE candidates, the caller learns the callee's IP addresses. The callee's server-reflexive address reveals a lot of information about the callee's location. In order to avoid tracking, implementations may wish to suppress the start of ICE negotiation until the callee has answered. In addition, either side may wish to hide their location from the other side entirely by forcing all traffic through a TURN server. In ordinary operation, the site learns the browser's IP address, though it may be hidden via mechanisms like Tor or a VPN. However, because sites can cause the browser to provide IP addresses, this provides a mechanism for sites to learn about the user's network environment even if the user is behind a VPN that masks their IP address. Implementations may wish to provide settings which suppress all non-VPN candidates if the user is on certain kinds of VPN, especially privacy-oriented systems such as Tor. See for additional information.

Communications Security Finally, we consider a problem familiar from the SIP world: communications security. For obvious reasons, it MUST be possible for the communicating parties to establish a channel which is secure against both message recovery and message modification. (See for more details.) This service must be provided for both data and voice/video. Ideally the same security mechanisms would be used for both types of content. Technology for providing this service (for instance, SRTP , DTLS , and DTLS-SRTP ) is well understood. However, we must examine this technology in the WebRTC context, where the threat model is somewhat different. In general, it is important to understand that unlike a conventional SIP proxy, the calling service (i.e., the Web server) controls not only the channel between the communicating endpoints but also the application running on the user's browser. While in principle it is possible for the browser to cut the calling service out of the loop and directly present trusted information (and perhaps get consent), practice in modern browsers is to avoid this whenever possible. "In‑flow" modal dialogs which require the user to consent to specific actions are particularly disfavored as human factors research indicates that unless they are made extremely invasive, users simply agree to them without actually consciously giving consent . Thus, nearly all the UI will necessarily be rendered by the browser but under control of the calling service. This likely includes the peer's identity information, which, after all, is only meaningful in the context of some calling service. This limitation does not mean that preventing attack by the calling service is completely hopeless. However, we need to distinguish between two classes of attack:

Retrospective compromise of calling service:: The calling service is non-malicious during a call but subsequently is compromised and wishes to attack an older call (often called a "passive attack").
During-call attack by calling service:: The calling service is compromised during the call it wishes to attack (often called an "active attack").

Providing security against the former type of attack is practical using the techniques discussed in . However, it is extremely difficult to prevent a trusted but malicious calling service from actively attacking a user's calls, either by mounting a Man-in-the-Middle (MITM) attack or by diverting them entirely. (Note that this attack applies equally to a network attacker if communications to the calling service are not secured.) We discuss some potential approaches in .

Protecting Against Retrospective Compromise In a retrospective attack, the calling service was uncompromised during the call, but an attacker subsequently wants to recover the content of the call. We assume that the attacker has access to the protected media stream as well as full control of the calling service. If the calling service has access to the traffic keying material (as in Security Descriptions (SDES) ), then retrospective attack is trivial. This form of attack is particularly serious in the Web context because it is standard practice in Web services to run extensive logging and monitoring. Thus, it is highly likely that if the traffic key is part of any HTTP request it will be logged somewhere and thus subject to subsequent compromise. It is this consideration that makes an automatic, public key-based key exchange mechanism imperative for WebRTC (this is a good idea for any communications security system), and this mechanism SHOULD provide Forward Secrecy (FS). The signaling channel/calling service can be used to authenticate this mechanism. In addition, if end-to-end keying is used, the system MUST NOT provide any APIs to either extract long-term keying material or to directly access any stored traffic keys. Otherwise, an attacker who subsequently compromised the calling service might be able to use those APIs to recover the traffic keys and thus compromise the traffic.

Protecting Against During-Call Attack Protecting against attacks during a call is a more difficult proposition. Even if the calling service cannot directly access keying material (as recommended in the previous section), it can simply mount a man-in-the-middle attack on the connection, telling Alice that she is calling Bob and Bob that he is calling Alice, while in fact the calling service is acting as a calling bridge and capturing all the traffic. Protecting against this form of attack requires positive authentication of the remote endpoint such as explicit out-of-band key verification (e.g., by a fingerprint) or a third-party identity service as described in .

Key Continuity One natural approach is to use "key continuity". While a malicious calling service can present any identity it chooses to the user, it cannot produce a private key that maps to a given public key. Thus, it is possible for the browser to note a given user's public key and generate an alarm whenever that user's key changes. The Secure Shell (SSH) protocol uses a similar technique. (Note that the need to avoid explicit user consent on every call precludes the browser requiring an immediate manual check of the peer's key.) Unfortunately, this sort of key continuity mechanism is far less useful in the WebRTC context. First, much of the virtue of WebRTC (and any Web application) is that it is not bound to a particular piece of client software. Thus, it will be not only possible but routine for a user to use multiple browsers on different computers that will of course have different keying material (Securely Available Credentials (SACRED) notwithstanding). Thus, users will frequently be alerted to key mismatches which are in fact completely legitimate, with the result that they are trained to simply click through them. As it is known that users routinely will click through far more dire warnings , it seems extremely unlikely that any key continuity mechanism will be effective rather than simply annoying. Moreover, it is trivial to bypass even this kind of mechanism. Recall that unlike the case of SSH, the browser never directly gets the peer's identity from the user. Rather, it is provided by the calling service. Even enabling a mechanism of this type would require an API to allow the calling service to tell the browser "this is a call to user X." All the calling service needs to do to avoid triggering a key continuity warning is to tell the browser that "this is a call to user Y" where Y is confusable with X. Even if the user actually checks the other side's name (which all available evidence indicates is unlikely), this would require (a) the browser to use the trusted UI to provide the name and (b) the user to not be fooled by similar appearing names.

Short Authentication Strings ZRTP uses a "Short Authentication String" (SAS) which is derived from the key agreement protocol. This SAS is designed to be compared by the users (e.g., read aloud over the voice channel or transmitted via an out-of-band channel) and if confirmed by both sides precludes MITM attack. The intention is that the SAS is used once and then key continuity (though with a different mechanism from that discussed above) is used thereafter. Unfortunately, the SAS does not offer a practical solution to the problem of a compromised calling service. "Voice cloning" systems, which mimic the voice of a given speaker are an active area of research and are already being used in real-world attacks . These attacks are likely to improve in future, especially in an environment where the user just wants to get on with the phone call. Thus, even if the SAS is effective today, it is likely not to be so for much longer. Additionally, it is unclear that users will actually use an SAS. As discussed above, the browser UI constraints preclude requiring the SAS exchange prior to completing the call and so it must be voluntary; at most the browser will provide some UI indicator that the SAS has not yet been checked. However, it is well known that when faced with optional security mechanisms, many users simply ignore them . Once users have checked the SAS once, key continuity is required to avoid them needing to check it on every call. However, this is problematic for reasons indicated in . In principle it is of course possible to render a different UI element to indicate that calls are using an unauthenticated set of keying material (recall that the attacker can just present a slightly different name so that the attack shows the same UI as a call to a new device or to someone you haven't called before), but as a practical matter, users simply ignore such indicators even in the rather more dire case of mixed content warnings.

Third-Party Identity The conventional approach to providing communications identity has of course been to have some third-party identity system (e.g., PKI) to authenticate the endpoints. Such mechanisms have proven to be too cumbersome for use by typical users (and nearly too cumbersome for administrators). However, a new generation of Web-based identity providers (BrowserID, Federated Google Login, Facebook Connect, OAuth , OpenID , WebFinger ) has been developed and use Web technologies to provide lightweight (from the user's perspective) third-party authenticated transactions. It is possible to use systems of this type to authenticate WebRTC calls, linking them to existing user notions of identity (e.g., Facebook adjacencies). Specifically, the third-party identity system is used to bind the user's identity to cryptographic keying material which is then used to authenticate the calling endpoints. Calls which are authenticated in this fashion are naturally resistant even to active MITM attack by the calling site. Note that there is one special case in which PKI-style certificates do provide a practical solution: calls from end users to large sites. For instance, if you are making a call to Amazon.com, then Amazon can easily get a certificate to authenticate their media traffic, just as they get one to authenticate their Web traffic. This does not provide additional security value in cases in which the calling site and the media peer are one and the same, but might be useful in cases in which third parties (e.g., ad networks or retailers) arrange for calls but do not participate in them.

Page Access to Media Identifying the identity of the far media endpoint is a necessary but not sufficient condition for providing media security. In WebRTC, media flows are rendered into HTML5 MediaStreams which can be manipulated by the calling site. Obviously, if the site can modify or view the media, then the user is not getting the level of assurance they would expect from being able to authenticate their peer. In many cases, this is acceptable because the user values site-based special effects over complete security from the site. However, there are also cases where users wish to know that the site cannot interfere. In order to facilitate that, it will be necessary to provide features whereby the site can verifiably give up access to the media streams. This verification must be possible both from the local side and the remote side. I.e., users must be able to verify that the person called has engaged a secure media mode (see ). In order to achieve this, it will be necessary to cryptographically bind an indication of the local media access policy into the cryptographic authentication procedures detailed in the previous sections. It should be noted that the use of this secure media mode is left to the discretion of the site. When such a mode is engaged, the browser will need to provide indicia to the user that the associated media has been authenticated as coming from the identified user. This allows WebRTC services that wish to claim end-to-end security to do so in a way that can be easily verified by the user. This model requires that the remote party's browser be included in the TCB, as described in .

Malicious Peers One class of attack that we do not generally try to prevent is malicious peers. For instance, no matter what confidentiality measures you employ the person you are talking to might record the call and publish it on the Internet. Similarly, we do not attempt to prevent them from using voice or video processing technology for hiding or changing their appearance. While technologies (Digital Rights Management (DRM), etc.) do exist to attempt to address these issues, they are generally not compatible with open systems and WebRTC does not address them. Similarly, we make no attempt to prevent prank calling or other unwanted calls. In general, this is in the scope of the calling site, though because WebRTC does offer some forms of strong authentication, that may be useful as part of a defense against such attacks.

Privacy Considerations

Correlation of Anonymous Calls While persistent endpoint identifiers can be a useful security feature (see ), they can also represent a privacy threat in settings where the user wishes to be anonymous. WebRTC provides a number of possible persistent identifiers such as DTLS certificates (if they are reused between connections) and RTCP CNAMEs (if generated according to rather than the privacy-preserving mode of ). In order to prevent this type of correlation, browsers need to provide mechanisms to reset these identifiers (e.g., with the same lifetime as cookies). Moreover, the API should provide mechanisms to allow sites intended for anonymous calling to force the minting of fresh identifiers. In addition, IP addresses can be a source of call linkage .

Browser Fingerprinting Any new set of API features adds a risk of browser fingerprinting, and WebRTC is no exception. Specifically, sites can use the presence or absence of specific devices as a browser fingerprint. In general, the API needs to be balanced between functionality and the incremental fingerprint risk. See .