- Possibilities and challenges of AI are rapidly evolving, and AI adoption is currently vastly outpacing AI security, leaving security as an afterthought.
- Standards are currently fragmented, leading to uncertainty and a slowdown in innovation. A cohesive standards landscape is required to enable speed and safe adoption.
- We currently have too many standards. There is a critical need to establish consensus, coordinate, and extract best practices across the existing standards.
- remain independent
- choose collaboration over competition
- choose lightweight coordination over procedures and committees
- choose standard landscape clarity over gains for individual organizations
- choose collective coordination over 1-on-1 or no coordination
- choose open coordination
- strive for being aligned on terminology and taxonomy, to prevent confusion
- strive for being open about their deliverables, plans, and roadmap, to facilitate collaboration, and prevent duplicate work
- avoid introducing yet another competing standard (referring to the famous XKCD comic) - which includes MOSAIC itself
- have a published general scope in which they operate as swimlanes. Scope-changes are announced timely
- strive for machine readable resources, as machines will increasingly be consumers of standards and standards meta data
- choose referring to existing work over producing work that is similar to what already exists
- don't overpromise: remain transparent in deliverables about the limitations of achieving security, and the reality of residual risks in AI
- act by the principles in the charter
- are actively represented by nominating at least one representative who is a subject matter expert that can participate in taxonomy/alignment discussions or meetings. Participants should have sufficient mandate or influence within their organization to make alignment meaningful (e.g. directly involved in maintaining its relevant resources).
- maintain transparency by keeping their member card current with public information about their in-scope work, including their scope, deliverables, and publicly announced plans. Members update the card when that information changes and review it at least twice a year. They also share announcements on the MOSAIC platform when particularly relevant for other members. Examples include scope changes and new initiatives that others may want to join.
- participate in discussions regularly on the MOSAIC platform, when relevant, including alignment on terminology and insights.
- In scope: The organization publishes or maintains standards, frameworks, guidance, controls, taxonomies, or other substantive material specifically addressing the security of AI systems.
- Established: The organization has published and maintained relevant AI-security material for at least 12–24 months.
- Recognized / adopted: Its work is demonstrably used or referenced outside its own organization, e.g. by standards bodies, regulators, government agencies, industry frameworks, academic publications, or multiple independent organizations.
- Organizational substance: The work is maintained by an identifiable organization or community with documented governance and named maintainers/editors, rather than by a single individual.
- Active: Relevant material has been updated, reviewed, or actively maintained within the preceding 12 months.
- In agreement and committed: The organization subscribes to the charter and is committed to fulfil the lightweight coordination requirements in it (see below).
- Are voted on and accepted if above criteria are met and 2/3 of casted votes are positive, provided existing participants get sufficient time to cast the vote (to be decided - for example 4 weeks).
The topic scope for MOSAIC is: protecting CIA for AI systems, where "I" includes the integrity of intended model behaviour.
That should not keep us from incorporating boundary aspects that are relevant, such as autonomy, transparency, excessive agency, model misuse, responsibility and accountability