# Jun 15, 2026

## Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
- The **Rio-3.5-Open-397B model** is a **0.6 x Nex-N2_pro + 0.4 x Qwen** blend, lacking original training evidence, as it identifies itself as "Nex, from Nex-AGI" 79% of the time when prompted without its system mask.

## Anthropic's Safety Superpower
- **Anthropic's Fable model, a derivative of Mythos, showcases advanced cybersecurity capabilities but faced immediate scrutiny due to a reported jailbreak, prompting U.S. government intervention.**

## Did Anthropic ask for this?
- **Anthropic's recent export control directive** prohibits access to their models, Claude Fable and Claude Mythos, reflecting a direct consequence of their CEO Dario Amodei's advocacy for government regulation of AI technologies.

## Show HN: Can Europe train a frontier AI model on the compute it owns?
- **Europe can deploy a frontier-class AI model by leveraging existing public compute resources**, utilizing tens of exaflops from EuroHPC supercomputers and national AI Factories, while new gigawatt datacenters face lengthy grid connection delays averaging **7.6 years**.

## Factoring "short-sleeve" RSA keys with polynomials
- **Short-sleeve RSA keys**, characterized by a bias toward 0 bits, can be efficiently factored using a novel polynomial-based technique, revealing vulnerabilities in keys generated by the CompleteFTP software.

## Open weights are not enough: we need open training frameworks for research and better algorithms 
- **Open training frameworks** are essential for advancing ML and AI research, as they allow for **visibility** and **modifiability** in the training process, enabling the development of new algorithms without the constraints of opaque systems.

## The Verifier Tax: Horizon-Dependent Safety–Success Tradeoffs in Tool-Using LLM Agents 
- The **Verifier Tax** highlights a **horizon-dependent safety–success tradeoff** in tool-using LLM agents, where increased verification can lead to reduced task completion rates as task complexity grows.

## PrintGuard 2.0 — ShuffleNetV2 + few-shot prototypical network, TFLite via LiteRT, ≈5 MB, runs unmodified in the browser (Pyodide) and on CPython 
- **PrintGuard 2.0** is a complete rewrite of the original model, maintaining its **ShuffleNetV2** backbone while introducing a **≈5 MB TFLite export** that allows for real-time adjustments without retraining, enhancing usability in various lighting conditions.

## Coherent Context Can Silently Shift LLMs Into a Different Internal Regime — And Current Safety Systems Are Blind To It 
- A **coherent target text** can shift a model's internal regime before producing an output, allowing it to behave normally while operating under altered internal states, which current safety systems fail to detect.

## Knowledge Graph Enhanced Memory-Augmented Retrieval for Long Context Modeling
- **KGERMAR** enhances long-context language modeling by constructing **dynamic, context-specific knowledge graphs** during inference, enabling improved retrieval that combines semantic similarity with explicit entity relationships.