Data privacy and security are critical to enterprise AI adoption – businesses need AI solutions that not only work for their specific needs, but keep their intellectual property (IP), customer data, and employee information secure. As part of our focus on real-world risks associated with AI technology and enterprise customers’ needs, Cohere has published our Secure AI Frontier Model Framework to describe our applied approach to safety and security.

This framework demonstrates how we develop our secure AI products and is part of several industry commitments Cohere has participated in, including those from the 2024 Seoul Summit in collaboration with the UK government.

Cohere’s approach: Practical risks, security, and privacy

We are not focused on speculative risks, but on the real-world, practical challenges that businesses and governments face today when using AI. Enterprise AI risks and mitigation strategies are distinct from AI integrated in consumer-facing chatbots or smartphone AI assistants that are available to the general public.

Customers across different industries and sectors use our AI models and applications at scale, for uses ranging from customer support and supply chain management to extracting and organizing financial information, analyzing legal agreements, answering questions about company policies, and summarizing text-based information.

Our customers have high standards and operate in industries where security and data privacy are paramount, like finance, healthcare, critical infrastructure, and the public sector. In addition to centering security with our comprehensive range of deployments like private clouds and on-premises, we focus on enterprises’ unique regulatory, safety, and security needs through our model development process.

Our Secure AI approach includes three core principles:

Evidenced Risks: We focus on real-world risks that are known, measured, or observable.
Assessed in Context: Risks are identified and assessed in the context of both Cohere’s and Cohere’s customers' specific use cases and contexts.
Holistically Managed: Safety and Security risk management is embedded by design into model and system development.

Our framework at a glance

The Cohere Secure AI Frontier Model Framework contains five components: (i) Risk Identification; (ii) Risk Assessment and Mitigation; (iii) Risk Assurance Mechanisms; (iv) Transparency; and (v) Research and External Stakeholder Engagement. Parts (i) to (iii) describe our risk management practices and how they are implemented throughout the AI lifecycle. Parts (iv) and (v) describe how we share information and best practices as well as support external research to continuously improve our own practices and advance the state-of-the-art in AI risk management.

Framework component	Our activities and practices
1. Risk Identification Understanding the type and nature of risks associated with our AI models and systems.	1a Identifying risks arising from model capabilities 1b Identifying risks to our systems 1c Understanding possible harms in context
2. Risk Assessment and Mitigation Applying measures to reduce risk throughout the lifecycle of our models and systems.	2a Defense-in-depth 2b Addressing harms 2c Secure AI by design
3. Risk Assurance Mechanisms Conducting tests and evaluations to verify and demonstrate that AI risks are appropriately mitigated.	3a Evaluations and benchmarking 3b Internal and third-party vulnerability testing 3c Customer-led testing and evaluation
4. Transparency Sharing information about our risk management practices and inviting feedback to ensure continuous improvement.	4a Documenting and sharing our practices 4b Continuous improvement
5. Research and External Stakeholder Engagement Supporting independent research and engaging externally to advance the science of AI risk management and contribute to industry standards.	5a Support for independent research 5b External engagement

Our full framework is available for download here. It is the first published version and will be updated as we continue to develop new best practices to advance the safety and security of our products.

Phil Blunsom