The report
Meta's coding-harness subscription terms make submitted development context eligible for model training by default. Under Section 3.3 of the Model API terms, Coding Harness Subscriptions use Contributor Services unless the account setting is switched to Standard Services. The designation applies regardless of the model selected through that subscription.
Contributor treatment matters because a coding assistant submits more than what the operator types. Meta's definition of Inputs includes material that the service is authorized to read through connectors, integrations and background context. A file excerpt or tool result retrieved by the agent can therefore become part of a request. Under Section 6.1, Meta may use Contributor Content for model training, development, evaluation and improvement. Under Section 5.1, requests to Standard Services are not used to train Meta Models.
Changing the setting does not rewrite earlier submissions
The terms say an account change to Standard applies prospectively. Meta retains its rights regarding Content already submitted under Contributor. They also specify that material previously submitted to Standard becomes Contributor Content if later included in a Contributor request, even as context rather than a direct user prompt. This is a concrete cross-request data-use boundary for systems that automatically assemble context from earlier conversations or repositories.
Section 6.2 bars sensitive, personal and confidential submissions to Contributor Services. The restriction is not evidence that the service can reliably identify confidential content arriving through a tool. A deployment handling protected source code therefore needs to check the service tier before handing the agent access to that material, not rely on a later prompt asking the model to withhold it.
The underlying Case distinguishes a documented contractual permission from verified use of any particular customer's repository in training. No such individual training event is established here. Retention and account-specific processing also require their own documentation, rather than being assumed from a promise that Standard content will not train models.