RELATED UPDATES

FMF Announces First-Of-Its-Kind Information-Sharing Agreement

The Frontier Model Forum (FMF) is pleased to share a progress update on our information-sharing efforts.

Last March the FMF announced a first-of-its-kind voluntary agreement that enables signatories to share information about vulnerabilities, threats, and concerning capabilities unique to frontier AI. 

The information-sharing function we have established addresses a critical need: facilitating greater exchange of information on AI-related risks to public safety and security, including information related to chemical, biological, radiological, and nuclear (CBRN) and advanced cyber threats.

What We’ve Accomplished 

Over the past year, the FMF has designed and implemented appropriate information-sharing protocols and policies for information related to the following:

• Vulnerabilities, weaknesses, and exploitable flaws that compromise the safety, security, or intended functionality of frontier AI models. Examples of what may be shared include information and artifacts related to jailbreaks, adversarial inputs, data poisoning, or other attempts to bypass model safeguards.
• Threats refer to attempts to gain unauthorized access or manipulate frontier AI models. Examples of what may be shared include information about potential threat actors or attack vectors, as well as cyber-threat indicators. 
• Capabilities of Concern: Frontier AI capabilities with the potential to cause large-scale societal harm. Examples of what may be shared include information about capabilities related to the development of CBRN threats, offensive cybersecurity attacks, and model autonomy. 

We are pleased to report that the member firms of the FMF have successfully shared the covered information above. From security vulnerabilities and jailbreak repositories to enhanced safety protocols, the information shared has supported firms’ efforts to improve the safety and security of their frontier AI models and systems. Our legal and technical infrastructure has supported the secure, confidential exchange of threat related information while protecting intellectual property and ensuring antitrust compliance.  

The information-sharing mechanism of the FMF has also become important to our other two mandates: developing industry best practices and advancing the science of AI safety. Most notably, the FMF has leveraged the information-sharing agreement to develop resources for mitigating cyber and biological risks from dual-use prompts, including emerging practice documents for model outputs and related evaluation datasets. We aim to share more about these resources later this year. 

Looking Ahead

We continue to explore ways to grow and scale our impact through information-sharing. Our upcoming priorities include: 

• Expanding participation in our existing information-sharing agreement. We intend to pilot voluntary information-sharing with non-FMF parties, starting with other frontier AI companies. 
• Building other trusted channels. As needed, we aim to establish additional secure, structured pathways for sharing threat-related information among industry, governments and relevant stakeholders.
• Sharing tools and resources. We aim to leverage our information-sharing function to develop and share a range of artifacts that benefit the broader AI safety and security ecosystem, but are too sensitive to release publicly given their associated information hazards. 
• Harmonizing frameworks. As related information-sharing pathways emerge, including those for incident reporting, we aim to develop practices and frameworks that facilitate interoperability where appropriate. 

When the FMF was founded, we were given a mandate to “establish trusted, secure mechanisms for sharing information among companies, governments and relevant stakeholders regarding AI safety and risks.” 

We remain committed to that founding mandate. With our core information-sharing mechanism in place, we aim to build on it in ways that strengthen the collective capacity of the frontier AI ecosystem to anticipate and mitigate large-scale risks.