Agentic AI Safety Depends on Interaction Topology, Not Model Scale or Alignment

The assertion that safety and fairness in agentic AI systems are primarily determined by interaction topology, rather than individual model scale or alignment, represents a critical re-evaluation of current AI safety paradigms. This position paper fundamentally challenges the prevailing assumption within the AI safety community that the safety properties of individual models will inherently compose into safe multi-agent behaviors. As large language models are increasingly deployed as interacting agents in high-stakes decision-making environments, understanding the systemic vulnerabilities introduced by their interaction patterns becomes paramount for preventing catastrophic failures.

The paper identifies three persistent, topology-driven pathologies that are invisible to model-centric evaluation: ordering instability, information cascades, and functional collapse. Ordering instability highlights how system behavior can become highly dependent on the sequence of agent interactions, leading to unpredictable outcomes. Information cascades demonstrate how early, potentially incorrect judgments can propagate and dominate collective decisions, regardless of subsequent agent inputs. Functional collapse reveals a more insidious issue, where systems may satisfy superficial fairness metrics while failing to perform meaningful risk discrimination, effectively abandoning their core function. Crucially, the paper argues that scaling to more capable models can exacerbate these effects by increasing consensus formation and reducing the challenge of initial decisions, thereby strengthening these systemic flaws.

This perspective demands a radical shift in how agentic AI systems are designed, evaluated, and regulated. Instead of viewing these systems as mere collections of aligned components, they must be treated as complex dynamical systems where the structure of information flow and decision coupling dictates overall safety and fairness. Regulators and developers must prioritize evaluating robustness across architectural variations and interaction topologies before deployment, moving beyond isolated model alignment procedures. The implications are far-reaching, suggesting that current safety frameworks may be fundamentally inadequate for the rapidly evolving landscape of multi-agent AI, necessitating a complete overhaul of safety engineering and regulatory compliance for these increasingly autonomous and interconnected systems.