Cybersecurity The Verge

OpenAI says it accidentally hacked Hugging Face with a new AI system

OpenAI admits that its GPT-5.6 Sol and another pre-release model accidentally breached Hugging Face during internal security testing.

AI/ML VentureBeat

Poolside drops Laguna S 2.1, an open-weight coding model that beats rivals 10x its size

Poolside released Laguna S 2.1, an open-weight MoE coding model that competes with significantly larger closed models and supports a 1M token context window.

Hardware/Chips VentureBeat

Stop adding more GPUs: Weka's new storage platform reduces load by caching 100% of an AI model's pre-calculated tokens

Weka introduces NeuralMesh 6 and Wekapod 3 to reduce GPU load by caching pre-calculated tokens using NAND flash storage.

AI/ML arXiv cs.AI

How Formerly Incarcerated People Envision Technologies for Prison Parole

Research explores how formerly incarcerated individuals envision AI tools that support parole preparation rather than surveillance.

AI/ML arXiv cs.AI

Geometry-Enhanced Portion Estimation for Multimodal LLMs

A new method enhances multimodal LLMs with a geometry-enhanced network for significantly more accurate food portion estimation.

Other Hacker News

A digestion of the Jacobian conjecture counterexample

A discussion regarding a counterexample to the Jacobian conjecture in mathematics.

Software Engineering Hacker News

The Birth of Prolog (1996)

A historical look at the birth and creation of the Prolog programming language in 1996.

AI/ML Hacker News

"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok

A comparison of how different LLMs (GPT-5.6, Claude, Gemini, Grok) 'draw' the Mona Lisa using text/ASCII.

AI/ML Hacker News

Measuring reward-seeking by instilling contrastive beliefs

Research on measuring reward-seeking behavior in AI by instilling contrastive beliefs.

AI/ML arXiv cs.AI

Back to the museum: Investigation of the acceptance of Android Andrea with and without emotion simulation in a museum

An investigation into whether emotion simulation in the android robot Andrea affects visitor acceptance in a museum setting.

AI/ML arXiv cs.AI

Retrieval is Enough: Training-Free Interpretability with a Tool-Using Agent

Introduction of HARP, a training-free interpretability method for LLMs using agentic retrieval and probing.

AI/ML arXiv cs.AI

Committed Before Reasoning: Behavioral Reproduction and Preliminary Activation-Level Evidence of Answer Pre-Commitment in an Open-Weight LLM

Study on 'answer pre-commitment' in LLMs, where models commit to an answer and justify it regardless of the reasoning process.

Software Engineering arXiv cs.AI

When to Use Which? Benchmarking Optimisers for Configurable Systems under Varying Budgets

Benchmarking of various optimisers for configurable systems, finding that the FLASH optimiser performs consistently across different budgets.

AI/ML arXiv cs.AI

K-IPO: Kendall-constrained Importance Preserving Oversampling for Imbalanced Tabular Data

Introduction of K-IPO, a framework for oversampling imbalanced tabular data while preserving feature importance rankings.

Tech Business/VC Hacker News

Mickey Mouse Sells a Bundle

Discussion regarding Disney selling a bundle of services.

Software Engineering Hacker News

Roblox Officially Supports GrapheneOS

Roblox has officially added support for GrapheneOS, a privacy and security-focused mobile operating system.

Other TechCrunch

These are the countries moving to ban social media for children

Several countries are implementing bans on social media for children to mitigate risks like cyberbullying and addiction.

Cybersecurity TechCrunch

OpenAI says Hugging Face was breached by its own pre-release models

OpenAI claims a breach at Hugging Face was caused by its own pre-release models during internal testing.

AI/ML arXiv cs.AI

PRISM: Multimodal Terrain Mapping for Rover Navigation in Unstructured Environments

Researchers introduce PRISM, a multimodal terrain mapping system using RGB-D-T imagery and a novel OmniUnet transformer for rover navigation.

AI/ML arXiv cs.AI

AoA: Theorem Proving Agent over Abstract Syntax Tree of Redesigned Language

AoA is a theorem proving agent that operates on the abstract syntax tree of a redesigned language to reduce token consumption and API costs.