When your CISO mentions “AI safety” within the subsequent board assembly, what precisely do they imply? Are they speaking about defending your AI methods from assaults? Utilizing AI to catch hackers? Stopping staff from leaking information to an unapproved AI service? Guaranteeing your AI doesn’t produce dangerous outputs?
The reply is likely to be “the entire above”; and that’s exactly the issue.
AI grew to become deeply embedded in enterprise operations. In consequence, the intersection of “AI” and “safety” has change into more and more complicated and complicated. The identical phrases are used to explain essentially completely different domains with distinct aims, resulting in miscommunication that may derail safety methods, misallocate assets, and go away important gaps in safety. We want a shared understanding and shared language.
Jason Lish (Cisco’s Chief Data Safety Officer) and Larry Lidz (Cisco’s VP of Software program Safety) co-authored this paper with me to assist deal with this problem head-on. Collectively, we introduce a five-domain taxonomy designed to convey readability to AI safety conversations throughout enterprise operations.
The Communication Problem
Contemplate this situation: your govt group asks you to current the corporate’s “AI safety technique” on the subsequent board assembly. And not using a widespread framework, every stakeholder might stroll into that dialog with a really completely different interpretation of what’s being requested. Is the board asking about:
- Defending your AI fashions from adversarial assaults?
- Utilizing AI to reinforce your risk detection?
- Stopping information leakage to exterior AI companies?
- Offering guardrails for AI output security?
- Guaranteeing regulatory compliance for AI methods?
- Defending in opposition to AI-enabled or AI-generated cyber threats? This ambiguity results in very actual organizational issues, together with:
- Miscommunication in govt and board discussions
- Misaligned vendor evaluations— evaluating apples to oranges
- Fragmented safety methods with harmful gaps
- Useful resource misallocation specializing in the unsuitable aims
And not using a shared framework, organizations wrestle to precisely assess dangers, assign accountability, and implement complete, coherent AI safety methods.
The 5 Domains of AI Safety
We suggest a framework that organizes the AI-security panorama into 5 clear, deliberately distinct domains. Every addresses completely different considerations, includes completely different risk actors, requires completely different controls, and usually falls beneath completely different organizational possession. The domains are:
- Securing AI
- AI for Safety
- AI Governance
- AI Security
- Accountable AI
Every area addresses a definite class of dangerous and is designed for use along side the others to create a complete AI technique.
These 5 domains don’t exist in isolation; they reinforce and rely on each other and should be deliberately aligned. Study extra about every area within the paper, which is meant as a place to begin for trade dialogue, not a prescriptive guidelines. Organizations are inspired to adapt and lengthen the taxonomy to their particular contexts whereas preserving the core distinctions between domains.
Framework Alignment
Simply because the NIST Cybersecurity Framework supplies a standard language to speak concerning the domains of cybersecurity whereas not eradicating the necessity for detailed cybersecurity framework reminiscent of NIST SP 800-53 and ISO 27001, this taxonomy will not be meant to work in isolation of extra detailed frameworks, however somewhat to supply widespread vocabulary throughout trade.
As such, the paper builds on Cisco’s Built-in AI Safety and Security Framework not too long ago launched by my colleague Amy Chang. It additionally aligns with established trade frameworks, such because the Coalition for Safe AI (CoSAI) Threat Map, MITRE ATLAS, and others.
The intersection of AI and safety will not be a single downside to unravel, however a constellation of distinct threat domains; every requiring completely different experience, controls, and organizational possession. By aligning with these domains with organizational context, organizations can:
- Talk exactly about AI safety considerations with out ambiguity
- Assess threat comprehensively throughout all related domains
- Assign accountability clearly to the precise groups
- Make investments strategically somewhat than reactively
