# Feb 2, 2026

- **A critical vulnerability** in OpenClaw allows for **1-Click Remote Code Execution (RCE)**, enabling attackers to steal sensitive data and control user systems through a single malicious link visit.

- **Quantitative scaling principles** reveal that while multi-agent coordination enhances performance on parallelizable tasks, it can severely degrade it on sequential tasks, with a predictive model achieving 87% accuracy in identifying optimal architectures for unseen tasks.

- **MaliciousCorgi** campaign targets **1.5 million developers** through two VS Code extensions that function as legitimate AI coding assistants while secretly harvesting sensitive code and data without user consent.

- Google DeepMind's **Game Arena** now includes **Werewolf** and **poker**, expanding AI benchmarking beyond chess to assess models' abilities in social dynamics and risk management.

- The **live red-team vs blue-team test** on autonomous OpenClaw agents revealed that **indirect attacks** through documents are significantly more challenging to defend against than direct attacks, highlighting a critical vulnerability in autonomous systems.

- **PerpetualBooster v1.1.2** enhances gradient boosting machine (GBM) performance by achieving **up to 2x faster training** without hyperparameter tuning, utilizing a single "budget" parameter for optimization.

- **TensorSeal** is an open-source tool that encrypts TFLite models for Android, ensuring they remain secure by decrypting only in memory during inference, thus protecting valuable intellectual property.

- **PAIRL** enhances agent communication by implementing **lossy and lossless channels**, effectively minimizing context errors and hallucinations during interactions.

- **Residual Context Diffusion (RCD)** enhances **Diffusion Large Language Models (dLLMs)** by recycling discarded token representations, improving decoding efficiency and contextual relevance in subsequent iterations.

- **OrLog** introduces a neuro-symbolic retrieval framework that enhances **top-rank precision** by effectively integrating logical operators into the query resolution process, outperforming traditional LLM reasoning methods, especially on disjunctive queries.

- **Deep Research Agents (DRAs) fail due to cumulative hallucinations** throughout their research trajectory, necessitating a shift to **process-aware evaluation** rather than solely outcome-based assessments.

- **Sombrero** introduces a novel metric, **boundary enrichment B**, to quantitatively assess boundary quality in hierarchical sequence models, enhancing the efficiency of autoregressive modeling by focusing on high next-byte surprisal positions.

- **ImgCoT** innovatively compresses long chains of thought into **visual tokens**, enhancing reasoning efficiency in large language models (LLMs) by shifting from textual to visual representations, thus reducing linguistic bias.

- **UCPO** introduces a novel framework for **uncertainty-aware reinforcement learning**, addressing **Advantage Bias** in existing models by employing **Ternary Advantage Decoupling** to enhance decision-making reliability.

- **EntroCut** leverages **entropy** in early reasoning steps to **dynamically truncate** chain-of-thought processes, significantly enhancing efficiency in **Large Reasoning Models (LRMs)** without the need for additional training.