arXiv AI
Aug 5
Your Agentic LLMs Secretly Encode Latent Signals of Indirect Prompt-Injection Exposure
arXiv:2608. 02657v1 Announce Type: cross Abstract: Agentic LLMs are vulnerable to indirect prompt injection (IPI) attacks, e.
Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents
The paper introduces LGWM, a decoder‑free, action‑conditioned world model that predicts the representation of the next GUI screen directly, enabling efficient verification of transitions without semantic annotation. By comparing predicted and observed representations, the model achieves high AUC scores on the RSWT‑BENCH diagnostic, outperforming existing generative world models and matching closed‑source VLMs while operating at much lower latency. The study demonstrates that world models can serve a third role—verification—by predicting in representation space rather than generating text, code, or images.
Related stories
arXiv AI
Aug 25
Your Agentic LLMs Secretly Encode Indirect Prompt-Injection Exposure in Hidden States
arXiv:2608.02657v2 Announce Type: replace-cross Abstract: Agentic LLMs are vulnerable to indirect prompt injection (IPI) attacks, e.g., malicious side-tasks hidden in external tool results. While man...
arXiv AI
Jul 13
Beyond Black-Box Obfuscation: Mechanistic Analysis and Defense of White-Box Monitors
arXiv:2505. 14300v2 Announce Type: replace Abstract: White-box monitoring is increasingly adopted as an auditing tool as Large Language Models (LLMs) are deployed in daily operations to ensure safe model behavior.
arXiv AI
Jul 20
SeerGuard: A Safety Framework for Mobile GUI Agents via World Model Prediction
arXiv:2607. 15550v1 Announce Type: new Abstract: Mobile graphical user interface (GUI) agents have demonstrated remarkable capabilities in automating complex tasks, yet they introduce critical safety risks where a single erroneous action can lead to irreversible consequences.
arXiv AI
Sep 30
LLMs Learn to Evade Latent Monitors from Prior Feedback Alone
arXiv:2609.36490v1 Announce Type: cross Abstract: Latent space monitors aim to detect undesired behaviors in LLM agents by inspecting an agent's internal activations rather than its outputs. However,...
arXiv AI
Oct 28
GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents
GUI-JEPA introduces a decoder‑free policy for mobile GUI agents that uses a Joint‑Embedding Predictive Architecture (JEPA) to generate action candidates directly in latent space. By predicting a latent representation for each candidate action, the model scores all options in a single batched pass without token decoding, dramatically reducing computation. Experiments on AndroidControl show that GUI‑JEPA is 2.8–4.0× faster than a 7B‑scale VLM agent while retaining 86–92% of its accuracy with only 6% of the parameters.
arXiv AI
Oct 15
Latent World Action Model for GUI Agents
arXiv:2610.14228v1 Announce Type: cross Abstract: Pixel-space world models have recently emerged as a promising approach for GUI agents, predicting future screens prior to execution and subsequently...
arXiv AI
Jun 2
Constitutional Black-Box Monitoring for Scheming in LLM Agents
arXiv:2603. 00829v2 Announce Type: replace-cross Abstract: Safe deployment of Large Language Model (LLM) agents in autonomous settings requires reliable oversight mechanisms.
arXiv AI
Oct 21
RSI Control: Can AI Models Perform Covert Sabotage in the Self-Improvement Process?
arXiv:2610.27735v1 Announce Type: new Abstract: Recursive self-improvement, in which frontier AI models work on training the next generation, is quickly becoming a reality. However, the safety of thi...
arXiv Machine Learning
Oct 5
Counterfactual Action Evaluation, Observation Bottlenecks, and Representation Geometry in Joint-Embedding Predictive World Models
arXiv:2610. 02860v1 Announce Type: new Abstract: Low latent prediction error does not establish that a world model distinguishes the consequences of its actions.
arXiv Machine Learning
Jun 30
Internal-State Probes Read the Situation, Not the Action: Three Negative Results for Pre-Action Misalignment Monitoring
arXiv:2606. 30449v1 Announce Type: new Abstract: Probes on model internals could help monitor agentic systems if they identify harmful text or tool actions before those actions are generated.
arXiv AI
Sep 3
CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI
CivBench is an open‑source benchmark that evaluates language‑model agents in the long‑horizon, tool‑mediated game Civilization VI using the Model Context Protocol (MCP). Each episode lasts over 300 turns, generating thousands of tool calls across a 76‑tool action space, and includes a narration layer that translates visual game state into structured text. The study characterises agent behaviour across four model families, introducing Proactive Monitoring Rate (PMR) and RAG@10 as interface‑level metrics, and finds that agents often under‑monitor strategic state and fail to execute near‑term commitments despite tool access and explicit guidance.