Intended Use

Primary Use Cases

1. Web task automation in productivity scenarios: filling forms, online shopping, booking travel and restaurants, information seeking, and other repetitive web workflows where automating clicks, typing, and navigation saves user time.
2. Web-agent role inside broader agentic systems: serving as the browser-acting sub-agent for orchestrators (for example, MagenticBrain or other planning models) that delegate web-bound subtasks.
3. Grounding model for screen-action prediction: predicting coordinates and actions on a screenshot when integrated as a vision-grounded action head inside a larger agentic stack.
4. Research and benchmarking on computer-use agents: studying web-agent behavior, evaluating safety controls, and developing new agentic capabilities in controlled sandboxed environments.

Out-of-Scope Use Cases

Responsible AI Considerations

• Unintended or unsafe browser actions: ambiguous user instructions, misleading UI elements, and unexpected on-screen content can lead to incorrect clicks, form submissions, or navigation. Fara1.5-9B is trained to pause at "critical points" (entering personal information, payment or shipping details, completing purchases or bookings, sending emails, submitting job applications, signing into accounts, making phone calls), but the safety property depends on the model correctly identifying these situations.
• Prompt injection from web content: harmful or adversarial content on visited pages can attempt to manipulate the model's behavior. Allow-lists, block-lists, and watch-mode monitoring are the primary mitigations.
• Quality of service: the model is trained primarily on English and is not intended for multilingual use; non-English content may experience worse performance. English varieties with less representation in the training data may also experience worse performance than standard American English.
• Representation harms and stereotypes: like other language models, Fara1.5-9B may over- or under-represent groups of people or reinforce demeaning or negative stereotypes despite safety post-training.
• Inappropriate or offensive content: may produce content unsuitable for sensitive deployment contexts without additional, use-case-specific mitigations.
• Information reliability: may generate nonsensical or fabricated content, misattribute sources, or be misled by deceptive or low-quality online content. Outputs require verification before being acted on.
• Data exposure: sensitive credentials (passwords, authentication tokens, payment data, personal identifiers) should never be passed to the model. Sandboxing, scoped permissions, and strict access controls are required to prevent leakage.

• Use MagenticLite where possible: the harness is built for safe Fara1.5-9B deployment and provides allow-lists, watch-mode, an immediate pause control, and sandboxing.
• Always run with human-in-the-loop: a human should monitor Fara1.5-9B's actions on the live web and be able to halt them immediately.
• Sandbox the browser: run Fara1.5-9B in a containerized environment (Docker, VM) that does not expose host files, environment variables, or saved credentials.
• Restrict network reach: use allow-lists or block-lists to limit which sites Fara1.5-9B can visit; this reduces exposure to prompt-injection attacks from adversarial pages.
• Never share sensitive credentials: passwords, payment data, and personal identifiers should not be entered into the model context.
• Layer safety services: use Azure AI Content Safety or equivalent guardrails on inputs and outputs.
• High-risk scenarios: do not use Fara1.5-9B for legal, medical, financial, or other regulated advice, nor in safety-critical workflows. Add application-layer safeguards appropriate to the deployment context.
• Misuse and misinformation: build feedback mechanisms and monitor for fraud, spam, or malware-style abuse patterns. Inform end-users they are interacting with an AI system.
• Allocation decisions: Fara1.5-9B is not suitable for scenarios with consequential impact on legal status or the allocation of resources or life opportunities (housing, employment, credit, and similar) without further assessment and additional debiasing.
• Verify outputs: treat model outputs as recommendations rather than authoritative decisions. Hallucination, misattribution, and adversarial manipulation by web content are all possible.
• Commercial deployment: this release is intended for research and development. Commercial use requires additional testing and validation in the deployment context.

Training Data

1. Synthetic trajectory data:
2. Grounding data: curated datasets for predicting actions and coordinates on the screen, combining screenshots, textual annotations, and bounding-box labels.
3. UI understanding data: visual question answering, captioning, and OCR over web-page screenshots collected by the data generation pipeline.
4. Safety and instruction-following data: refusal datasets containing harmful task statements that the model is trained to refuse rather than act on. This subset is text-only.
5. Public web and reasoning datasets: open-source datasets used alongside the synthetic corpus to broaden coverage.