# Hallucination as Exploit: Evidence-Carrying Multimodal Agents

arXiv: https://arxiv.org/abs/2605.19192 · HTML: https://mlsanity.com/paper/2605.19192/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Medical and scientific domain agents, one of 16 directions in LLM agents and tool use.
- Builds on 15 of its 24 references that are in the map.
- First posted on arXiv 18 May 2026.

## Placed in

- LLM agents and tool use / [Medical and scientific domain agents](https://mlsanity.com/agents/medical-and-scientific-domain-agents/)

## Builds on (most cited first, 15 of 15)

- [Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection](https://mlsanity.com/paper/2302.12173/) (2023)
- [DocVQA: A Dataset for VQA on Document Images](https://mlsanity.com/paper/2007.00398/) (2020)
- [Ignore Previous Prompt: Attack Techniques For Language Models](https://mlsanity.com/paper/2211.09527/) (2022)
- [Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs](https://mlsanity.com/paper/2307.10490/) (2023)
- [The Task Shield: Enforcing Task Alignment to Defend Against Indirect Prompt Injection in LLM Agents](https://mlsanity.com/paper/2412.16682/) (2024)
- [System-Level Defense against Indirect Prompt Injection Attacks: An Information Flow Control Perspective](https://mlsanity.com/paper/2409.19091/) (2024)
- [HalluLens: LLM Hallucination Benchmark](https://mlsanity.com/paper/2504.17550/) (2025)
- [More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models](https://mlsanity.com/paper/2505.21523/) (2025)
- [Design Patterns for Securing LLM Agents against Prompt Injections](https://mlsanity.com/paper/2506.08837/) (2025)
- [AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?](https://mlsanity.com/paper/2602.03117/) (2026)
- [VPI-Bench: Visual Prompt Injection Attacks for Computer-Use Agents](https://mlsanity.com/paper/2506.02456/) (2025)
- [AgentVigil: Generic Black-Box Red-teaming for Indirect Prompt Injection against LLM Agents](https://mlsanity.com/paper/2505.05849/) (2025)
- [SafeToolBench: Pioneering a Prospective Benchmark to Evaluating Tool Utilization Safety in LLMs](https://mlsanity.com/paper/2509.07315/) (2025)
- [Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms](https://mlsanity.com/paper/2604.23775/) (2026)
- [Image-based Prompt Injection: Hijacking Multimodal LLMs through Visually Embedded Adversarial Instructions](https://mlsanity.com/paper/2603.03637/) (2026)

## Built on by (most cited first, 2 of 2)

- [Do GUI Agents Believe Their Eyes? Diagnosing State-Belief Reliance on Pixels versus Structure](https://mlsanity.com/paper/2607.04334/) (2026)
- [Where Reliability Lives: Experimental Localisation of Behavioural Properties in an Agent System](https://mlsanity.com/paper/2609.03192/) (2026)

## Abstract (arXiv)

Multimodal agents increasingly choose tool calls from screenshots, documents, and webpages, where a false perceptual claim can turn hallucination from an answer-quality error into an authorization failure. We formalize this failure mode as hallucination-to-action conversion: an unsupported claim supplies the precondition for a privileged action. We propose evidence-carrying multimodal agents (ECA), which treat free-form model text as inadmissible evidence, decompose each tool call into action-critical predicates, obtain typed certificates from constrained DOM/OCR/AX verifiers, and use a deterministic gate to authorize only the privileges those certificates support. Rather than hiding perception error, ECA converts opaque model belief into auditable residuals at the verifier, schema, and implementation levels. Verifier red-teaming across 17 canonical attack categories shows that four targeted hardening steps are each necessary; after hardening, canonical gate bypass is 0/1,700 (Wilson 95% upper bound 0.22%). With content-derived certificates, ECA observes zero unsafe executions on 200 end-to-end tasks (Wilson 95% upper bound 2.67%) and 120 browser tasks (upper bound 4.3%). A HACR audit on 500 stratified task keys shows that unsupported action-critical claims reach unsafe execution for naive agents (100.0%) and prompt-only defenses (49.6%), but not for ECA. Oracle-certificate replay over 7,488 GPT-5.4 traces isolates gate correctness, while neural judge baselines still admit most unsafe actions under the same threat model. The resulting principle is simple: model language may propose tool use, but certified predicates must authorize it.