paulononato/backstage
1
0
Fork 0
Code Issues Pull Requests Actions 47 Packages Projects Releases Wiki Activity

Merge pull request #24640 from backstage/rugvip/threat-model

Browse Source 
Threat Model updates
This commit is contained in:
Patrik Oldsberg
2024-05-09 14:07:03 +02:00
committed by GitHub
parent 012c081663 b6582bc2cf
commit 3c47d9b755
1 changed files with 23 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/overview/threat-model.md
+23 -14
View File
@@ -14,21 +14,28 @@ The Backstage trust model is divided into three groups with different trust leve
An **internal user** is an authenticated user that generally belongs to the organization of a particular Backstage deployment. These users are trusted to the extent that they are not expected to compromise the availability of Backstage, but they are not trusted to not compromise data confidentiality or integrity.
An **integrator** is a user responsible for configuring and maintaining an instance of Backstage. Integrators are fully trusted, since they operate the system and database and therefore have root access to the host system. Additional measures can be taken by adopters of Backstage in order to restrict or observe the access of this group, but that falls outside of the current scope of Backstage.
An **operator** is a user responsible for configuring and maintaining an instance of Backstage. Operators are fully trusted, since they operate the system and database and therefore have root access to the host system. Additional measures can be taken by adopters of Backstage in order to restrict or observe the access of this group, but that falls outside of the current scope of Backstage.
Another group of de facto integrators is internal and external code contributors. When installing Backstage plugins you should vet them just like any other package from an external source. While it’s possible to limit the impact of for example a supply chain attack by splitting the deployment into separate services with different plugins, the Backstage project itself does not aim to prevent these kinds of attacks or in any other way sandbox or limit the access of plugins.
A **builder** is an internal or external code contributor and end up having a similar level of access as operators. When installing Backstage plugins you should vet them just like any other package from an external source. While it’s possible to limit the impact of for example a supply chain attack by splitting the deployment into separate services with different plugins, the Backstage project itself does not aim to prevent these kinds of attacks or in any other way sandbox or limit the access of plugins.
An **external user** is a user that does not belong to the other two groups, for example a malicious actor outside of the organization. The security model of Backstage currently assumes that this group does not have any direct access to Backstage, and it is the responsibility of each adopter of Backstage to make sure this is the case.
## Integrator Responsibilities
## Operator Responsibilities
As an integrator of Backstage you yourself are responsible for protecting your Backstage installation from external and unauthorized access. The sign-in system in Backstage does not exist to limit access, only to inform the system of the identity of the user. There are some plugins that have more fine-grained access control through the permissions system, but the primary purpose of that system is to restrict access to resources for internal users rather than Backstage as a whole. A common and recommended way to protect a Backstage deployment from unauthorized access is to deploy it behind an authenticating proxy such as AWS’s ALB, GCP’s IAP, or Cloudflare Access.
:::info
This section assumes that you are using the
[new backend system](../backend-system/index.md) and at least Backstage release [version 1.24](../releases/v1.24.0.md). Before that Backstage did not come with built-in protection against unauthorized access and you were required to deploy it in a protected environment.
:::
Other responsibilities include protecting the integrity of configuration files as it may otherwise be possible to introduce vulnerable configurations, as well as the confidentiality of configured secrets related to Backstage as these typically include authentication details to third party systems.
Backstage is primarily designed to be deployed in a protected environment rather than being exposed to the public internet. From a confidentiality and integrity perspective, Backstage is designed to protect against unauthorized access to data and to ensure that data is not tampered with. However, Backstage does not provide more than rudimentary protection against denial of service attacks, and it is the responsibility of the operator to ensure that the Backstage deployment is protected against such attacks. A common and recommended way to protect a Backstage deployment from unauthorized access is to deploy it behind an authenticating proxy such as AWS’s ALB, GCP’s IAP, or Cloudflare Access.
The integrator is ultimately responsible for auditing usage of internal and external plugins as these run on the host system and have access to configuration and secrets. When installing plugins from sources like NPM, you should vet these in the same way that you would vet any other package installed from that source.
Users that are signed-in in to Backstage generally have full access to all information and actions. If more fine-grained control is required, the [permissions system](../permissions/overview.md) should be enabled and configured to restrict access as necessary.
The integrator is also responsible for maintaining the resolved NPM dependencies of their Backstage project. This involves ensuring that `yarn.lock` receives updated versions of packages that have vulnerabilities, when those fixed versions are in range of what the Backstage packages request in their respective `package.json` files. This is commonly done by employing automated tooling such as [Dependabot](https://dependabot.com/), [Snyk](https://snyk.io/), and/or [Renovate](https://docs.renovatebot.com/) on your own repository. When fixed versions exist that are _not_ in range of what Backstage packages request, or when larger operations such as switching out an entire dependency for another one is required, maintainers collaborate with contributors to try to address those dependency declarations in the main project as soon as possible.
An operator is responsible for protecting the integrity of configuration files as it may otherwise be possible to introduce vulnerable configurations, as well as the confidentiality of configured secrets related to Backstage as these typically include authentication details to third party systems.
The operator is ultimately responsible for auditing usage of internal and external plugins as these run on the host system and have access to configuration and secrets. When installing plugins from sources like NPM, you should vet these in the same way that you would vet any other package installed from that source.
The operator is also responsible for maintaining the resolved NPM dependencies of their Backstage project. This involves ensuring that `yarn.lock` receives updated versions of packages that have vulnerabilities, when those fixed versions are in range of what the Backstage packages request in their respective `package.json` files. This is commonly done by employing automated tooling such as [Dependabot](https://dependabot.com/), [Snyk](https://snyk.io/), and/or [Renovate](https://docs.renovatebot.com/) on your own repository. When fixed versions exist that are _not_ in range of what Backstage packages request, or when larger operations such as switching out an entire dependency for another one is required, maintainers collaborate with contributors to try to address those dependency declarations in the main project as soon as possible.
## Common Backend Configuration
@@ -44,21 +51,23 @@ Note that the `UrlReader` system operates with a service context and is not inte
Backstage provides authentication of users through the `auth` plugin, which primarily acts as an authorization server for different OAuth 2.0 provider integrations. These integrations can both serve the purpose of signing users into Backstage, as well as providing delegated access to external resources, and are all subject to the common concerns of implementing secure OAuth 2.0 authorization servers. All auth provider integrations are disabled by default, and need to be enabled through configuration in order to be used. For each Backstage installation it is recommended to only enable the minimal set of providers that are in use by that instance.
It is not within scope of the `auth` backend to protect against unauthorized access, that is something that needs to be handled at a deployment level. See the [Integrator Responsibilities](#integrator-responsibilities) section for more information.
In order to use an auth provider to sign in users into Backstage, it needs to be configured with an [Identity resolver](https://backstage.io/docs/auth/identity-resolver), which is a custom callback implemented in code. The identity resolver is a sensitive part of configuring Backstage and it is important that it always resolves user identities correctly, based on information provided by the authentication provider. There are a number of built-in identity resolvers that can simplify configuration, and it is important that these all resolve users in a secure way, regardless of how they are used.
Backstage also supports authentication through an authenticating reverse proxy such as [AWS ALB](https://aws.amazon.com/elasticloadbalancing/application-load-balancer/), where the user identity is read from the incoming proxied decorated request. The following proxy auth providers verify the signature of incoming requests, and are therefore safe to deploy with direct access by users: `awsAlb`, `cfAccess`, and `gcpIap`. Providers like `oauth2Proxy` do not verify the incoming request and can therefore be spoofed by a malicious internal user to supply the `auth` backend with forged identity information. It’s therefore highly recommended to restrict access to the `oauth2Proxy` endpoints, or use a different provider.
As part of signing in with an identity resolver, a Backstage Token is issued containing the resolved user identity. The tokens are asymmetrically signed JSON Web Tokens, with the public keys available to any service that wishes to verify a token. The signing keys are rotated continuously and are unique to each installation of Backstage, meaning that Backstage Tokens are not shared across installations. The token contains claims for the user identity and ownership information, which can be used to determine what Backstage resources are owned by that user or group. It is important that this token can not be forged outside of the `auth` plugin, with the exception of other plugins deployed in the same backend service or sharing the same database. For a high-security deployment, the `auth` backend should therefore be deployed in a separate service with its own database.
The token is used to prove the identity of the user within the Backstage system, and is used throughout Backstage plugins to control access. It is important that the ownership resolution logic is consistent across the entire Backstage ecosystem, with no possibility of misinterpreting the ownership information.
For cross-backend communication, the Backstage Token is typically forwarded or, in strict backend-to-backend communication without a user party, the backend itself issues a service token based on a pre-shared secret which is then validated on the receiving end. There are no unique service identities tied to these tokens at this point, meaning the tokens can be used across all services in a Backstage installation. This is something that we aim to improve in the future.
One of the claims in a user token is the User Identity Proof or `uip`. This is an additional signature of the token that allows for offline token transformation. By replacing the original signature with the `uip` the token is still proof of a user identity, but it no longer acts as a full access token and will be rejected by most plugin endpoints. Plugins can explicitly allow use of this limited token where required, but this should only be used when necessary when a full token is not available, and ideally just for read-only access. Use-cases for limited users tokens include cookie authentication of static assets, storage of user identity proofs in a database, and similar.
Backstage also supports authentication through a proxy where the user identity is read from the incoming request from the proxy, which has been decorated by an authenticating reverse proxy such as [AWS ALB](https://aws.amazon.com/elasticloadbalancing/application-load-balancer/). The following proxy auth providers verify the signature of incoming requests, and are therefore safe to deploy with direct access by users: `awsAlb`, `cfAccess`, and `gcpIap`. Providers like `oauth2Proxy` does not verify the incoming request and can therefore be spoofed by a malicious internal user to supply the `auth` backend with forged identity information. It’s therefore highly recommended to restrict access to the `oauth2Proxy` endpoints, or use a different provider.
The communication across backend plugins uses a similar authentication scheme to the user authentication. Each backend plugin generates and publishes its own set of keys that it uses to sign its tokens, and the public keys are shared with all other plugins for verification. The expected location of each plugin's published JWKS is determined by the `DiscoveryService` implementation in the backend, which means that it is vital for any custom implementation of that service to be careful with user input. The tokens signed by each plugin contain both the source and target plugin ID, which means that the token can not be reused to access other plugins.
When forwarding a user identity in a call across backend plugins only the limited user token with `uip` is used, wrapped in a new service token that is signed by the calling plugin. This means that the receiving plugin can trust the user identity, but it is not able to make further calls on behalf of the user except for with the plugins that it is authorized to call. That is except for any endpoints in other plugins that accept limited user tokens, which is a reason to avoid accepting them when possible.
## Catalog
Integrators should configure [catalog rules](https://backstage.io/docs/features/software-catalog/configuration#catalog-rules) to limit the allowed entity kinds that users can define. In general it is best to restrict definition of User, Group, and Template entities so that internal users cannot register additional ones. Template entities define actions that are executed on the backend hosts, and while the goal is for these actions to be secure regardless of input, it is still a more sensitive context and it is recommended that you protect it with additional checks. It is very important to not allow registration of User and Group entities if you ingest and rely on these as organizational data in your catalog. Doing so could otherwise open up for the ability to impersonate users and confuse group membership information. You should always ingest organizational data using a statically configured catalog location or an entity provider reading from a trusted source. The entities emitted directly by an entity provider are always trusted and rules are not applied to them, but any entities produced further down the chain are still subject to the rules.
Operators should configure [catalog rules](https://backstage.io/docs/features/software-catalog/configuration#catalog-rules) to limit the allowed entity kinds that users can define. In general it is best to restrict definition of User, Group, and Template entities so that internal users cannot register additional ones. Template entities define actions that are executed on the backend hosts, and while the goal is for these actions to be secure regardless of input, it is still a more sensitive context and it is recommended that you protect it with additional checks. It is very important to not allow registration of User and Group entities if you ingest and rely on these as organizational data in your catalog. Doing so could otherwise open up for the ability to impersonate users and confuse group membership information. You should always ingest organizational data using a statically configured catalog location or an entity provider reading from a trusted source. The entities emitted directly by an entity provider are always trusted and rules are not applied to them, but any entities produced further down the chain are still subject to the rules.
The Catalog does not aim to protect against resource exhaustion attacks in its default setup. If you need to prevent your internal users from being able to register large amounts of entities, then it is recommended to disable entity registration and use a different approach for discovering entities. One way to mitigate any resource exhaustion attacks is to only allow the catalog to read from trusted SCM sources that have an audit trail. Catalog currently lacks limits for entity hierarchy depth and entity size, which we hope to address in the future.
@@ -68,11 +77,11 @@ By default all internal users are allowed to create and delete entities. If this
By default, Scaffolding jobs execute directly on the host machine, including any actions defined in the template. Because the Scaffolder templates are considered a more sensitive area it is recommended to control access to create and update templates to trusted parties. Template execution is intended to be secure regardless of input, but we still recommend this additional layer of protection. The string templating is executed in a [node VM sandbox](https://github.com/laverdet/isolated-vm) to mitigate the possibility of remote code execution attacks.
The Scaffolder often has elevated permissions to for example create repositories in a Github organization. The integrator should therefore be cautious of Scaffolder Templates that for example delete or update existing resources as the user input is typically user defined and can therefore delete or modify resources maliciously or by mistake.
The Scaffolder often has elevated permissions to for example create repositories in a Github organization. The operator should therefore be cautious of Scaffolder Templates that for example delete or update existing resources as the user input is typically user defined and can therefore delete or modify resources maliciously or by mistake.
One strategy that allows you to reduce the access that the Scaffolder service has is to rely on user credentials when executing actions. For example, a GitHub App integration could be configured with read-only permissions, with a separate user OAuth token used to create repositories. This requires that your users have access to create repositories in the first place.
The integrator should audit installed scaffolding actions just like any other plugin package. It is also important to verify that installed actions fall in line with your own security requirements, as some actions might be intended for more relaxed environments.
The operator should audit installed scaffolding actions just like any other plugin package. It is also important to verify that installed actions fall in line with your own security requirements, as some actions might be intended for more relaxed environments.
By default all internal users are allowed to execute templates in the scaffolder. If this does not fit your organization's needs it is recommended to enable and configure the [permission](https://backstage.io/docs/permissions/overview) system to restrict these operations.