Research map / Adversarial robustness and security / Attacks on text and graphs

EnvTrap: Revealing the Environment-Only Attack Surface in Embodied AI via Consequence-Blind Action Execution

NeurIPS 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: LLM jailbreak and injection defense

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies environment-only perturbations as an attack surface for embodied AI agents, requiring no model or instruction access, exposing consequence-blind action execution.

Problem

Prior attacks on embodied AI focus on semantic or instruction-level manipulations (prompts, adversarial images, action commands); the paper targets attacks that only modify the physical environment without access to the model or instructions.

Approach

Introduces environment-only perturbations as a new attack surface for embodied agents, which exploit consequence-blind action execution: the agent executes actions without anticipating environmental consequences.

Closest papers by meaning

• AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions (CVPR 2026 · similarity 0.90)
• Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-Based Decision-Making Systems (ICLR 2025 · similarity 0.88)
• Embodied Active Defense: Leveraging Recurrent Feedback to Counter Adversarial Patches (ICLR 2024 · similarity 0.86)
• AdvEDM: Fine-grained Adversarial Attack against VLM-based Embodied Agents (NeurIPS 2025 · similarity 0.86)
• IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks (AAAI 2026 · similarity 0.86)
• TRAP: Targeted Redirecting of Agentic Preferences (NeurIPS 2025 · similarity 0.86)