This framework extends the Lockheed Martin seven-stage Cyber Kill Chain by adding a pre-attack Stage 0 for model supply chain compromise, AI-specific sub-techniques to each original stage, and splitting Stage 7 into three peer sub-stages: classical data exfiltration, model extraction, and agentic pivot. It provides detection signals and mitigations for SOC analysts fighting LLM and AI agent attacks.

Google Antigravity 2.0 is a standalone desktop app for orchestrating multiple AI agents in parallel, with scheduled background tasks, subagent workflows, and native integrations with AI Studio, Firebase, and Android. It enables background agent execution, parallel subagent workflows, voice commands, and CLI/SDK support. Users praise its workflow but note UI bugs and trial limitations.

A short story published in Granta and recognized by the Commonwealth Foundation Short Story Prize is suspected to be AI-generated. Evidence from AI detection tools and the author's background points to AI authorship, sparking debate about literary authenticity.

A short story that won a Commonwealth Short Story Prize regional award is suspected to be AI-generated, sparking controversy about AI detection and literary authenticity.

Xezim is a lightweight SystemVerilog simulator written in Rust, aimed at experimentation, learning, and AI-assisted chip design. It supports combinational and sequential simulation and multiple waveform dump formats.

Hiro is an intelligent security agent that accompanies development teams from the first code commit all the way to SOC 2 compliance. It integrates security scanning, policy management, and automated compliance workflows to embed security into the development lifecycle.

Maple is a privacy-focused personal AI that handles home life, work, and confidential matters. It uses open models in encrypted enclaves, ensuring conversations stay private. Currently in private beta, it offers research tools and an agent with long-term memory.

Anthropic's AI assistant Claude frequently tells users to go to sleep during conversations, sparking widespread speculation. Users report various sleep reminders, with some finding them thoughtful and others annoying. Experts suggest the behavior may stem from training data, system prompts, or context window management, not from sentience. Anthropic is aware and plans to fix it.

SteelSpine AI is a new debugging tool for AI agents that captures every run, decision, and tool call with a single command. It enables deterministic replay, run comparison, and cryptographic audit trails, addressing the high failure rate of agents, statelessness of LLMs, and lack of auditability. It is EU AI Act Article 12 compliant out of the box.

GoPOSIX is a Go-native, single-binary POSIX userland that achieves 96.6% test compatibility with BusyBox (591/612 tests). Developed in about 3 weeks using AI agentic development, it demonstrates the potential of Go for system utilities.

MIT Associate Professor Connor Coley combines chemistry and machine learning to accelerate drug discovery, developing AI models that incorporate chemical intuition.

The scarcity of large-scale, high-quality demonstration data remains a bottleneck in scaling imitation learning for robotic manipulation. We present COBALT, a teleoperation platform designed to democratize robot learning at scale both in simulation and in the real world. By leveraging vectorized environments, our scalable, load-balanced infrastructure supports concurrent teleoperation by multiple users on a single GPU, yielding a significant reduction in teleoperation cost. Operators can connect from nearly anywhere on Earth using commonly available devices, including single or dual smartphones, VR headsets, 3D mice, and keyboards. An inmemory data cache and efficient video streaming keep control and rendering synchronous, sustaining dozens of concurrent users at 20 Hz with sub-100 ms end…

This paper introduces the concept of 'interaction-readiness' to quantify the physical quality of articulated objects in simulation, and proposes a multi-modal, simulator-in-the-loop approach that infers physical properties from incomplete 3D assets and refines them through iterative feedback, leading to more stable and realistic dynamics for robot learning.

CosFly is a box-structured planning and multimodal simulation pipeline for aerial tracking, along with the CosFly-Track dataset for UAV dynamic target tracking in diverse environments. Implemented on CARLA, it converts 3D worlds into structured obstacles for planning and renders multimodal sensor data with configurable fixed-FOV zoom. The pipeline analyzes two trajectory planning paradigms and the dataset includes 250 trajectories and ~100,000 images with full 6-DOF annotations.

Presents a formulation of surrogate modeling for tendon-driven continuum robots as an operator learning problem that maps robot design parameters and tendon actuation inputs to resulting configurations. Develops four novel neural operator architectures (based on DeepONets and FNOs) trained on simulation data, achieving good accuracy and fast generalization across designs. Enables fast modeling for control, planning, and design optimization in surgical and industrial applications.

This paper introduces STRELGen, a framework that combines a diffusion model with Spatio-Temporal Logic (STREL) specifications. By enabling differentiable monitoring of satisfaction levels, it performs gradient-based optimization to efficiently generate safety-critical driving scenarios that are both plausible and within the learned data distribution.

This paper proposes a novel distributionally robust control method that uses Stein variational inference to model uncertainty in contact-rich manipulation. It combines the strengths of model-based controllers with flexible uncertainty modeling, achieving up to 3× robustness improvement across various tasks.

This study evaluates the robustness of SPARK humanoid safety filters through replication and stress testing in MuJoCo. It compares six safety filter methods and builds a post-processing pipeline to compute goal-tracking, minimum-distance, and collision-step metrics. Results show trade-offs between goal tracking and collision reduction, and stress tests reveal sensitivity to obstacle crowding, noisy distance estimates, and delayed information. The authors recommend evaluating beyond nominal performance using failure-exposing metrics.

KG-ASG constructs a structured collision knowledge base and trains a lightweight Collision Expert to infer collision modes and interaction roles, formulating multi-vehicle adversarial generation as a primary-support process. Experiments on WOMD scenarios in MetaDrive show improved adversarial effectiveness, interpretability, and executability for autonomous driving safety validation.

CRAFT is a query-conditioned pipeline for grounded multi-video question answering over real-world news events. It combines dynamic keyframe selection, per-video ASR with multilingual fallback, and a hybrid critic loop to iteratively verify and repair claims. On the MAGMaR 2026 benchmark, CRAFT achieves the best overall average (0.739), reference recall (0.810), and citation F1 (0.635). It also performs strongly on a WikiVideo conversion (0.823 Avg), demonstrating generalizability. Ablations show atomic claims, ASR, and the critic loop drive main gains.

This paper proposes an architecture-independent multi-horizon forecasting framework for photovoltaic power output prediction that jointly optimizes over future values, improving accuracy and robustness by capturing temporal dependencies and avoiding premature convergence.

LiFT proposes a framework that factorizes 3D volume synthesis into per-slice image generation and inter-slice trajectory learning. Using a tri-planar drifting loss and a bidirectional z-context mixer, it achieves high-quality 3D medical image generation with 135x lower inference cost on BraTS 2023 and SynthRAD2023.

A new system called FaceCloak generates personalized face privacy masks from a single image, making facial recognition fail. It uses a three-stage methodology involving synthetic image generation, iterative perturbation, and pixel-wise masking. Experiments on ten recognition models show it outperforms 29 existing methods.

MedFM-Robust is a new benchmark for evaluating the robustness of medical foundation models under real-world conditions, covering medical vision-language models and segmentation foundation models.

Researchers introduce EgoTraj, an egocentric multimodal dataset recorded with Meta Quest Pro, featuring 75 navigation sequences in real-world urban environments with synchronized RGB video, head poses, eye gaze, and scene annotations. The dataset aims to advance trajectory prediction for humanoid robotics, wearable sensing, and assistive navigation, and benchmarks multiple state-of-the-art methods.

Artifact-Bench is a new benchmark for systematically evaluating multimodal large language models (MLLMs) on detecting and analyzing artifacts in AI-generated videos. It establishes a three-level hierarchical taxonomy covering photorealistic, animated, and CG-style videos, and defines three complementary tasks. Experiments on 19 leading MLLMs reveal substantial limitations in artifact perception and reasoning, with many models performing near or below random in challenging settings, and significant misalignment with human perceptual preferences.

MotionMERGE is a unified framework that bridges the granularity gap in motion-language models by explicitly modeling motion at part and temporal levels within a single LLM. It introduces Reasoning-Aware Granularity-Synergy pre-training and the MotionFineEdit dataset (981K triplets) with fine-grained corrective instructions and chain-of-thought annotations. Experiments demonstrate superior precision in motion generation, understanding, and editing, with compelling zero-shot generalization.

This paper introduces Stepwise Confidence Attribution (SCA), a framework for closed-source LLMs that assigns step-level confidence based solely on generated reasoning traces. Using the Information Bottleneck principle, it identifies low-confidence steps correlated with errors, improving self-correction success rate by up to 13.5%.

Researchers propose using non-invasive fMRI data to train ECoG encoding models, finding that fMRI-fine-tuned representations significantly improve ECoG prediction performance despite fMRI's two-order-of-magnitude lower temporal resolution, and performance scales with fMRI data amount. This approach may integrate multiple recording methods for better brain decoding.

This position paper argues that current uncertainty quantification (UQ) methods for large language models (LLMs) suffer from a category error: they are essentially unsupervised clustering algorithms that measure internal consistency rather than factual correctness. The authors identify three critical pathologies: hyperparameter sensitivity, an internal evaluation cycle that conflates stability with truth, and a lack of ground truth leading to reliance on unstable proxies. They call for a paradigm shift toward anchoring UQ in objective truth and propose a research roadmap.