The New Baseline for AI Safety
AI is now not an experimental functionality or a back-office automation software: it’s changing into a core operational layer inside trendy enterprises. The tempo of adoption is breathtaking. But, based on Cisco’s 2025 AI Readiness Index, solely 29 % of firms consider they’re adequately geared up to defend in opposition to AI threats and solely 33 % have a proper change-management plan for guiding accountable adoption.
Executives and leaders more and more discover themselves in a troubling place: they perceive cybersecurity, however AI safety feels overseas. People, organizations, and governments can’t adequately comprehend or reply to the implications of such quickly evolving expertise and the threats that ensue: organizations are deploying techniques whose habits evolves, whose modes of failure usually are not absolutely understood, and whose interactions with their setting are dynamic and generally unpredictable.
Cisco’s Built-in AI Safety and Security Framework (additionally referred to on this weblog as “AI Safety Framework”) provides a essentially completely different method. It represents one of many first holistic makes an attempt to categorise, combine, and operationalize the total vary of AI dangers, from adversarial threats, content material security failures, mannequin and provide chain compromise, agentic behaviors and ecosystem dangers (e.g., orchestration abuse, multi-agent collusion), and organizational governance. This vendor-agnostic framework offers a construction for understanding how trendy AI techniques fail, how adversaries exploit them, and the way organizations can construct defenses that evolve alongside functionality developments.
A Fragmented Panorama—and the Want for Integration
For years, organizations that tried to safe AI pieced collectively steering from disparate sources. MITRE ATLAS helped outline adversarial techniques in machine studying techniques. NIST’s Adversarial Machine Studying taxonomy described assault primitives. OWASP printed Prime 10 lists for LLM and agentic dangers. Frontier AI labs like Google, OpenAI, and Anthropic shared inner security practices and rules. But every of those efforts centered on a selected slice of the chance panorama, providing items of the puzzle however cease wanting offering a unified, end-to-end understanding of AI danger.
What has been lacking is a cohesive mannequin—one which seamlessly spans security and safety, runtime and provide chain, mannequin habits and system habits, enter manipulation and dangerous outputs. Cisco’s evaluation makes the hole clear: no present framework covers content material harms, agentic dangers, provide chain threats, multimodal vulnerabilities, and lifecycle-level publicity with the completeness wanted for enterprise-grade deployment. The true world doesn’t section these domains, and adversaries actually don’t both.
Evaluation of protection throughout AI safety taxonomies and frameworks
A New Paradigm for Understanding AI Danger
AI safety and security dangers current very actual issues for organizations. Taken collectively, AI safety and AI security kind complementary dimensions of a unified danger framework: one involved with defending AI techniques from threats, and the opposite with making certain that their habits stays aligned with human values and ethics. Treating these domains in tandem can allow organizations to construct AI techniques that aren’t solely sturdy and dependable, but additionally accountable and worthy of belief.
We outline them as:
- AI safety: the self-discipline of making certain AI accountability and defending AI techniques from unauthorized use, availability assaults, and integrity compromise throughout the AI lifecycle.
- AI security: serving to guarantee AI techniques behave ethically, reliably, pretty, transparently, and in alignment with human values.
Cisco’s Built-in AI Safety and Security Framework is constructed upon 5 design components that distinguish it from prior taxonomic efforts and embody an evolving AI menace panorama: the combination of AI threats and content material harms, AI improvement lifecycle consciousness, multi-agent coordination, multimodality, and audience-aware utility.
(1) Integration of threats and harms: One core innovation of Cisco’s framework is its recognition that AI safety and AI security are inseparable. Adversaries exploit vulnerabilities throughout each domains, and oftentimes, hyperlink content material manipulation with technical exploits to realize their goals. A safety assault, akin to injecting malicious directions or corrupting coaching knowledge, usually culminates in a security failure, akin to producing dangerous content material, leaking confidential data, or producing undesirable or dangerous outputs.
Conventional approaches have handled security and safety as parallel tracks. Our AI Safety Framework makes an attempt to replicate the fact of contemporary AI techniques: the place adversarial habits, supposed and unintended system habits, and person hurt are interconnected. The AI Safety Framework’s taxonomy brings these components right into a single construction that organizations can use to grasp danger holistically and construct defenses that tackle each the mechanism of assault and the ensuing influence.
(2) AI lifecycle consciousness: One other defining function of the AI Safety Framework is its anchor within the full AI lifecycle. Safety issues throughout knowledge assortment and preprocessing differ from these throughout mannequin coaching, deployment and integration, software use, or runtime operation. Vulnerabilities which can be irrelevant throughout mannequin improvement might change into essential as soon as the mannequin features entry to tooling or interacts with different brokers. Our AI Safety Framework follows the mannequin throughout this whole journey, making it clear the place completely different classes of danger emerge and the way they might evolve, and permitting organizations to implement defense-in-depth methods that account for a way dangers evolve as AI techniques progress from improvement to manufacturing.
(3) Multi-agent orchestration: The AI Safety Framework can even account for the dangers that emerge when AI techniques work collectively, encompassing orchestration patterns, inter-agent communication protocols, shared reminiscence architectures, and collaborative decision-making processes. Our taxonomy accounts for related dangers that emerge in techniques with autonomous planning capabilities (brokers), exterior software entry (MCP1), persistent reminiscence, and multi-agent collaboration—threats that might be invisible to frameworks designed for earlier generations of AI expertise.
(4) Multimodality issues: The AI Safety Framework additionally displays the fact that AI is more and more multimodal. Threats can emerge from textual content prompts, audio instructions, maliciously constructed photographs, manipulated video, corrupted code snippets, and even embedded alerts in sensor knowledge. As we proceed to analysis how multimodal threats can manifest, treating these pathways constantly is crucial, particularly as organizations undertake multimodal techniques in robotics and autonomous automobile deployments, buyer expertise platforms, and real-time monitoring environments.
(5) An audience-aware safety compass: Lastly, the framework is deliberately designed for a number of audiences. Executives can function on the degree of attacker goals: broad classes of danger that map on to enterprise publicity, regulatory issues, and reputational influence. Safety leaders can deal with methods, whereas engineers and researchers can dive deeper into subtechniques. Drilling down even additional, AI pink groups and menace intelligence groups can construct, check, and consider procedures. All of those teams can share a single conceptual mannequin, creating alignment that has been lacking from the trade.
The AI Safety Framework offers groups with a shared language and psychological mannequin for understanding the menace panorama past particular person mannequin architectures. The framework contains the supporting infrastructure, complicated provide chains, organizational insurance policies, and human-in-the-loop interactions that collectively decide safety outcomes. This permits clearer communication between AI builders, AI end-users, enterprise capabilities, safety practitioners, and governance and compliance entities.
Contained in the AI Safety Framework: A Unified Taxonomy of AI Threats
A vital part of the AI Safety Framework is the underlying taxonomy of AI threats that’s structured into 4 layers: goals (the “why” behind assaults), methods (the “how”), subtechniques (particular variants of “how”), and procedures (real-world implementations). This hierarchy creates a logical, traceable pathway from high-level motivations to detailed implementation.
The framework identifies nineteen attacker goals, starting from objective hijacking and jailbreaks to communication compromise, knowledge privateness violations, privilege escalation, dangerous content material era, and cyber-physical manipulation. These goals map on to noticed patterns and threats, to vulnerabilities organizations are encountering as they scale AI adoption, and at last prolong to areas which can be technically possible, although not but noticed exterior of a analysis setting. Every goal turns into a lens by way of which executives and leaders can perceive their publicity: which enterprise capabilities might be impacted, which regulatory obligations is likely to be triggered, and which techniques require heightened monitoring.
Strategies and subtechniques present the specificity crucial for operational groups. These embody over 150 methods and subtechniques akin to immediate injections (each direct and oblique), jailbreaks, multi-agent manipulation, reminiscence corruption, provide chain tampering, environment-aware evasion, software exploitation, and dozens extra. The richness of this layer displays the complexity of contemporary AI ecosystems. A single malicious immediate might propagate throughout brokers, instruments, reminiscence shops, and APIs; a single compromised dependency might introduce unobserved backdoors into mannequin weights; or a single cascaded failure might trigger a whole multi-agent workflow to diverge from its supposed objective.
Screenshot of the AI Safety Framework’s Taxonomy Navigator
The protection taxonomy embedded inside the framework is equally sturdy. It contains twenty-five classes of dangerous content material, starting from cybersecurity misuse to security and content material harms to mental property compromise and privateness assaults. This breadth acknowledges that many AI failures are emergent behaviors that may nonetheless trigger real-world hurt. A unified taxonomy ensures that organizations can consider each malicious inputs and dangerous outputs by way of a coherent lens.
Alongside that vein, there are further mannequin context protocol (MCP), agentic, and provide chain menace taxonomies embedded inside the AI Safety Framework. Protocols like MCP and A2A govern how LLMs interpret instruments, prompts, metadata, and execution environments, and when these parts are tampered with, impersonated, or misused, benign agent operations will be redirected towards malicious targets. The MCP taxonomy (which presently covers 14 menace sorts) and our A2A taxonomy (which presently covers 17 menace sorts) are each standalone assets which can be additionally built-in into AI Protection and in our open supply instruments: MCP Scanner and A2A Scanner. Lastly, provide chain danger can be a core dimension of lifecycle-aware AI safety. We’ve developed a taxonomy that covers 22 distinct threats and is equally built-in into AI Protection, our companions in mannequin safety, and different instruments we’re growing for the open supply group.
Cisco’s Built-in AI Safety and Security Framework provides some of the full, forward-looking approaches out there in the present day. At a time when AI is redefining industries, that readability shouldn’t be merely beneficial—it’s important. This framework can be built-in into Cisco AI Protection, the place threats are recognized with related indicators and mitigation methods. Navigate our Built-in AI Safety and Security Framework in the present day. We stay up for working with the group to deepen the notice and strengthen defenses in opposition to this novel ecosystem of AI threats.
