Protecting the System, Stranding the User: The Limits of Multi-Agent Authorization

Gurpriya Bhatia, Clifford NeumanarXiv 2026 · arXiv

TL;DR

Current multi-agent security models are great at protecting systems but terrible at protecting users. This paper argues that "successful" security interventions often leave humans stranded—for example, by aborting a transaction after a non-refundable deposit is paid—and proposes a new architectural synthesis that binds recovery actions directly to authorization tokens.

What Problem Does This Paper Address?

In the burgeoning world of AI agents, we often delegate complex tasks (like booking an apartment) to a chain of specialized agents. Current security research focuses on system integrity: ensuring an agent doesn't overstep its permissions.

The authors identify a "blind spot": what happens to the human when the security system works exactly as intended? If a security monitor detects a privacy violation halfway through a workflow and kills the process, the system is "safe," but the user might be left with a forfeited deposit, an incomplete application, and a generic error message. The paper tackles the disconnect between deterministic security (did the agent follow the rules?) and human outcomes (did the user get what they wanted without losing money?).

Methods

This is a position paper that critiques the current "emerging consensus" in agentic security. The authors analyze existing frameworks (like the Sagas pattern, Macaroons, and Contextual Integrity) and propose four structural shifts to move toward "Human-Centered Authorization":

  1. Decoupled Reporting: Separating deterministic "rule-conformance" (Yes/No) from "intent-fidelity" (a probabilistic score).
  2. Evidence-Bound Handoffs: Requiring agents to pass a cryptographic summary of why they trust a request when crossing company boundaries, rather than just passing a "naked" trust score.
  3. Token-Bound Sagas: Moving the "undo" logic (compensating actions) from the high-level planning layer down into the low-level authorization token itself.
  4. Reputation Feedback Loops: Creating a system where automated recoveries penalize the initiating agent’s reputation to prevent "moral hazard."

Datasets

Not discussed in the paper. As a theoretical and architectural position paper, it does not perform empirical testing on specific datasets. Instead, it builds on established computer science theorems (like Rice’s Theorem) and existing security standards (like AuthZEN and Verifiable Intent).

Key Insights and Ideas

  • The Illusion of Verified Intent: The authors argue that "verifying intent" is a category error. Because a human's internal state is unobservable, intent has no formal ground truth. Systems should stop claiming they have "verified" intent and instead report an openly probabilistic estimate.
  • The Boundary Problem: A trust score calculated by your personal assistant means nothing to a third-party bank agent. Without the underlying evidence, a trust score is just a "naked number" that violates the principle of Contextual Integrity.
  • Security as a "Successful Failure": If a system halts a workflow but leaves the user in a worse position than if the violation had occurred, the authors argue this is a systemic failure, not a security success.
  • Moral Hazard in AI: If we make automated recovery (rollbacks) too "frictionless," developers might become lazy, relying on the "undo" button rather than building robust intent-alignment checks in the first place.

Key Results

The paper concludes with a proposed Architectural Synthesis. Rather than abandoning current tools like cryptographically signed intent or reference monitors, the authors suggest augmenting them with:

  • Risk-Anchored Thresholds: Security gates that scale their strictness based on how "irreversible" an action is (e.g., sending an email vs. spending $1,000).
  • Actionable Summaries: Instead of forensic logs, the system must generate human-readable summaries of what was completed, what was undone, and what is currently "stranded" or irreversible.
  • The Sagas Pattern in Tokens: By embedding the "remedy" (e.g., "if this fails, release the funds") into the capability token, the authority to act and the authority to fix a failure travel together across agent boundaries.

Limitations and Future Work

The authors candidly identify several "Unresolved Theoretical Limits":

  • The Calibration Paradox: How do you tune a security threshold for "intent" if you can never truly measure the ground truth of what a user wanted?
  • The Compression Limit: Can we summarize the "evidence" for trust enough to pass it between agents without leaking the user's entire private conversation history?
  • Physical Irreversibility: Some things cannot be "undone" by a protocol—like an inappropriately read email or a market price change.
  • Legal Liability: In a multi-agent chain spanning three different companies, who is legally responsible for executing (or paying for) a recovery action?

Summary updated

Abstract

Current multi-agent authorization architectures prioritize system integrity over human outcomes. While recent models successfully prevent privilege escalation, they externalize the cost of failure onto the human operator. Re-centering authorization around human impact exposes four structural vulnerabilities in current paradigms. First, calling an agent's understanding of user intent verified is misleading. Intent has no formal definition, so no check against it can be definitive. Systems should instead report an openly probabilistic estimate and scale their caution with the risk of the action. Second, we argue that dynamic trust scores do not survive multi-agent handoffs; transmitting algorithmic judgment across a company boundary without underlying evidence risks violating Contextual Integrity. Third, treating a safely halted workflow as a successful refusal is a category error if it strands the user; compensating actions must be bound directly to the capability token, not only to the planning layer. Finally, we hypothesize that frictionless automated recovery introduces a moral hazard, disincentivizing developers from building robust initial intent checks. We conclude by proposing an architectural synthesis to bridge the gap between deterministic security and human outcomes, alongside the fundamental theoretical limits that remain unresolved.
BibTeX
@misc{bhatia2026protecting,
  author = {Gurpriya Bhatia and Clifford Neuman},
  title = {{Protecting the System, Stranding the User: The Limits of Multi-Agent Authorization}},
  year = {2026},
  eprint = {2611.19598},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19598},
}
Figure