TLDR: This research paper introduces an 11-layer failure stack to identify vulnerabilities in conventional, generative, and agentic AI systems, ranging from hardware to agentic reasoning. It also develops ‘awareness mapping,’ a framework to assess organizational recognition of these risks, linking it to a five-level maturity scale. The paper highlights how reliability concerns evolve across different AI paradigms and advocates for a proactive, layered approach integrated with Dependability-Centred Asset Management (DCAM) for trustworthy AI deployment.
Artificial intelligence systems are becoming increasingly integral to critical sectors like transportation, energy, healthcare, and manufacturing. While conventional AI systems have long faced reliability challenges, the emergence of generative AI and agentic AI introduces new and complex vulnerabilities. A recent research paper, From Failure Modes to Reliability Awareness in Generative and Agentic AI System, by Janet (Jing) Lin and Liangwei Zhang, offers a comprehensive framework to understand and address these evolving reliability concerns.
The paper introduces an innovative 11-layer failure stack, a structured framework designed to identify vulnerabilities across the entire spectrum of AI systems. These layers range from the very foundational elements like hardware and power, through core intelligence components such as models and data, to operational aspects like applications and monitoring, and finally to the advanced agentic layer involving reasoning and multi-agent coordination. This layered approach highlights that failures rarely occur in isolation; instead, they often propagate across layers, leading to cascading effects with significant systemic consequences.
Reliability, as the paper explains, is not a static property but a dynamic and expanding concept. For conventional AI, reliability focuses on consistent predictions and classifications, with common failures including data drift, model instability, and hardware faults. Generative AI, which produces open-ended outputs like text or images, expands these concerns to include hallucinations, factual errors, biased content, and potential misuse. Agentic AI, with its autonomy, planning, and multi-agent interactions, further deepens the challenge, introducing risks such as goal misalignment, flawed reasoning, and emergent conflicts between agents.
The 11-Layer Failure Stack Explained
The 11 layers are broadly categorized into four domains:
- Foundational Layers: Hardware, Power & Energy, System Software, and AI Frameworks. These form the physical and computational base. Failures here can include overheating, voltage instability, kernel panics, or dependency conflicts.
- Core Intelligence Layers: Models and Data. These are the learning and knowledge base of AI. Vulnerabilities here involve overfitting, adversarial attacks, hallucinations in models, or data drift, noise, and mislabeling.
- Operational Layers: Applications, Execution, Monitoring, and Learning. These cover deployment and adaptation. Issues can range from poor application integration, latency spikes during execution, silent monitoring failures, to concept drift or biased retraining data in learning.
- Agentic Layer: AI Agent. This is where autonomy and decision-making unfold. Failures here include goal misalignment, emergent conflicts, and opaque decision-making.
The paper emphasizes that while all layers are susceptible to failures, the prominence of certain risks shifts depending on the AI paradigm. For conventional AI, data, models, and applications are key. For generative AI, the focus moves to model hallucinations, data bias from web-scale corpora, and robust monitoring for content safety. Agentic AI brings learning, execution, and the agent layer itself to the forefront, dealing with lifelong adaptation and complex multi-agent coordination.
Also Read:
- Charting a Sustainable Path for Artificial Intelligence: A Deep Dive into Green AI Practices
- Advancing Urban Planning with AI: Beyond Prediction to Transparent Reasoning
Awareness Mapping: A Path to Reliability Maturity
Beyond technical safeguards, the paper introduces “awareness mapping” – a maturity-oriented framework that quantifies how well individuals and organizations recognize reliability risks across the AI stack. This framework operationalizes the 11-layer failure stack into specific reliability studies, allowing organizations to score their awareness. The scores are then mapped onto a five-level maturity scale, ranging from “Unaware” (Level I) to “Comprehensive Cross-Layer Reliability” (Level V).
Empirical insights from practitioner surveys revealed a consistent trend: most organizations exhibit fragmented awareness, focusing on surface-level failures while overlooking deeper, more consequential risks. Awareness mapping serves as both a diagnostic tool to identify blind spots and a roadmap for guiding improvements in AI governance, training, monitoring, and lifecycle management.
Finally, the research links awareness mapping to Dependability-Centred Asset Management (DCAM), positioning AI systems as critical organizational assets whose dependability must be managed proactively. By integrating awareness into DCAM, organizations can move from reactive responses to proactive strategies, ensuring trustworthy and sustainable AI deployment in mission-critical domains.
