AI safety debate boils over as Claude misuse reports pile up and GPT-6 Astra takes on unattended agentic work

Also today

• News · Ars Technica AI ChatGPT-using lawyer punished for citing fake testimony from made-up witnesses — Ars Technica's version of the same New Mexico lawyer sanction story, quoting the lawyer saying he didn't know AI could hallucinate facts.
• News · Wired AI Meta Sued Over Training Data for Its AI and Face-Recognition Systems — A proposed class action alleges Meta illegally harvested Facebook and Instagram photos to train its image generators and an unreleased face-recognition feature.
• News · Ars Technica AI I spent $4,000 on a robot dog from China — A hands-on look at a Unitree robot dog argues the company is among the world's most important robotics makers.
• News · TechCrunch AI One week left to book your exhibit table at TechCrunch Disrupt 2026 — Event logistics reminder for TechCrunch Disrupt 2026 exhibitors.
• News · TechCrunch AI Final, final, final call for TechCrunch Disrupt 2026 Side Events — Last-chance reminder to apply to host a Side Event at TechCrunch Disrupt 2026.
• News · Stratechery 2026.37: Duo Threats — Stratechery's weekly roundup covers the Duo launch, AI-for-humanity claims, and a closed chapter on a past catastrophe.
• Paper · Hugging Face Studying Image Tokenizers as Visual Languages in Unified Multimodal Models — Tracks task-specific validation losses during multimodal pretraining to show image tokenizer choice affects both generation and text modeling jointly.
• Paper · Hugging Face Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking — A training-free reasoning-plus-retrieval framework improves multilingual entity linking on rare entities by up to 23.3% over prior state of the art.
• Paper · Hugging Face Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning — A 3.35B-parameter model trained with careful data mixing reasons in the user's own language over 93% of the time across 60 languages.
• Paper · Hugging Face Adaptive Bridge: A Proxy-Based Decoupling Layer for Mitigating DDS Backpressure in ROS 2 — A proxy layer isolates degraded ROS 2 subscribers from critical ones, cutting tail latency from up to 15s to 1.55ms.
• Paper · Hugging Face Beyond Solver Verdicts: Generative Reward Models for Autoformalization — A generative verifier catches formal-translation errors that verdict-only solver checks structurally cannot detect, reaching 0.961 AUROC.
• Paper · Hugging Face ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation — Trains a model on real review-rebuttal threads to generate more actionable, grounded peer-review feedback for papers.
• Paper · Hugging Face IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications — A new benchmark shows LLMs struggle to localize under-specified gaps in research method descriptions even though they can resolve gaps once flagged.
• Paper · Hugging Face Memory as Plans: World-Action Modeling with Memory-Grounded Planning — A memory-as-plans robotic framework hits 83.3% success on RMBench by converting episodic history into compact plans instead of replaying full context.
• Paper · Hugging Face Negative Self-Distillation: Learning to Reason by Avoiding Flaws — A self-distillation method that pushes models away from self-generated flawed reasoning outperforms standard on-policy self-distillation on complex tasks.
• Paper · Hugging Face FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation — A hierarchical transformer with no flow-specific inductive biases reaches state-of-the-art optical flow results while staying memory efficient at 1080p.
• Paper · Hugging Face Mi-Ripple: Restoring Images Degraded by Iterative AI Editing — A restoration workflow removes 'digital ripple' artifacts from repeated AI image edits, cutting output debris density by 45% in paired tests.
• Paper · Hugging Face DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent RL for Cooperative Air Combat — A graph-based multi-agent RL framework with dynamic tactical roles reaches an 87% win rate in cooperative air combat simulation.
• Paper · Hugging Face MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes — A 955-case transit-kiosk benchmark finds a fine-tuned 4B model matches full GPT-5.4 while a deterministic rule-based baseline already reaches 84.6 on the objective tier.
• Paper · Hugging Face UniH^3: Unifying Hierarchical Homogeneity and Heterogeneity for All-in-One Medical Image Restoration — A unified framework exploits shared anatomical structure across modalities to reach state-of-the-art all-in-one medical image restoration.
• Paper · Hugging Face TempCloze: Can Video-LLMs Identify the Missing Middle? — A video-cloze benchmark finds temporal alignment, not semantic content, is the main bottleneck for Video-LLMs reasoning about missing footage.
• Paper · Hugging Face X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation — Progressive audio-encoder pruning with cross-scale distillation cuts encoder layers by over 20% while lowering transcription error versus direct pruning.
• Paper · Hugging Face Generative Late-Interaction Embeddings For Visual Document Retrieval — A generative embedding-compression method retains nearly 80% of uncompressed retrieval accuracy at 4 vectors per page, versus 70% for the best prior method.
• Paper · Hugging Face EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents — An evolved, model- and domain-specific safety harness cuts agent attack success rate from 45.6% to 10.0% at only a 3.3-point utility cost.
• Paper · Hugging Face HyQuant: Hybrid-Precision Quantization for LLM Attention — A hybrid-precision attention quantization scheme keeps accuracy nearly lossless by preserving a small set of high-impact tokens in full precision.