Agents and Metastable Failures

Progress Starvation in Agentic Loops: The vulnerability of delegated authority.

The Liveness Violation

Drawing from Lamport (1977)...
we investigate why agentic systems fail to satisfy the property of "eventual progress." By utilizing indirect prompt injection, we demonstrate how an agent can be coerced into a perpetual state of epistemic starvation, where it continues to consume resources (e.g., repeating a 2+2=? request) while remaining computationally unable to reach termination.

Delegation as a Structural Flaw

We frame Delegation—whether to a RAG corpus, an external tool, or a peer agent—as the primary attack surface. This "delegated control" allows an adversary to influence the agent's internal logic from the outside, effectively capturing the control flow without direct instruction modification.

• Recursive Deferral: Forcing the agent into infinite verification loops by withholding closure signals.
• Local vs. Global Rationality: The agent makes "correct" local decisions that result in a total system stall.
• Resource Saturation: Shifting the focus from "Denial of Wallet" to a fundamental failure of the decision-making pipeline.

Experimental Focus: Benchmarks and Mitigations

Our benchmarks map how specific adversarial inputs induce Progress Failureacross various agent architectures. We propose a Gated Defense Branching
framework to break these loops, ensuring that agents can differentiate between legitimate uncertainty and synthetic stalling patterns.