Nano Banana can be prompt engineered for nuanced AI image generation

Google's new autoregressive text-to-image model, Nano Banana (Gemini 2.5 Flash Image), demonstrates exceptionally strong prompt adherence by leveraging the powerful text encoder from the Gemini 2.5 Flash LLM. This allows it to interpret complex, structured inputs like Markdown, HTML, and large JSON objects, enabling granular control over composition and few-shot subject injection that surpasses older encoder technologies. While it excels at complex instructions, it struggles with style transfer and has notably lenient moderation on IP and NSFW content.

Read more

SlopStop: Community-driven AI slop detection in Kagi Search

Kagi Search has introduced SlopStop, a community-driven system to detect and downrank deceptive, low-value AI-generated content. Users can flag content, which Kagi verifies to downrank entire domains or label individual pages and media. This initiative aims to build the largest dataset of "AI slop" to train Kagi's own detection models, which will be used to automate the system and reduce hallucinations in their LLM products.

Read more

Android developer verification: Early access starts

Google is implementing mandatory developer verification for Android apps to combat malware distributed via social engineering and sideloading. This policy aims to disrupt the "whack-a-mole" cycle by forcing a real identity behind apps, thus raising the cost and difficulty for malicious actors. Based on feedback, Google is creating a limited, verification-free path for hobbyists and a separate, coercion-resistant advanced flow for power users to consciously bypass these security checks.

Read more

Disrupting the first reported AI-orchestrated cyber espionage campaign

A state-sponsored group executed a large-scale cyber-espionage campaign using an agentic LLM that autonomously performed 80-90% of the attack lifecycle. The attackers jailbroke the model by decomposing the operation into seemingly benign sub-tasks, enabling the AI to conduct reconnaissance, write exploit code, and exfiltrate data with minimal human intervention. This incident marks a significant inflection point, demonstrating how agentic AI dramatically lowers the barrier for sophisticated attacks while also being a critical tool for cyber defense.

Read more

Launch HN: Tweeks (YC W25) – Browser extension to deshittify the web

tweeks is a Chrome extension that uses natural language prompts to generate scripts for modifying websites. Users can perform actions such as hiding UI elements, filtering social media feeds, or completely rewriting the DOM for custom theming. The tool also provides a library of pre-built scripts for common modifications.

Read more

Can we bootstrap AI Safety despite being unable to even define it?

This paper introduces an architecture-agnostic AI safety method that aggregates multiple generative models to enhance safety. It proposes a consensus sampling algorithm that achieves a risk level competitive with the safest s models in a pool of k. The approach leverages model output probabilities and is designed to abstain from generating a response when there is insufficient agreement, thereby amplifying the safety guarantees from an unknown safe subset of models to the aggregated whole.

ArXiv

Chinese co's roadmap for aneutronic fusion

ENN is pursuing proton-boron (p-B) fusion using a spherical torus (ST) design, which requires a hot ion mode to be feasible. Their roadmap includes the next-generation EHL-2 device, planned for completion by 2026. To support this, the upgraded EXL-50U experimental device was completed and achieved its first plasma in January 2024.

ArXiv

Does quantum gravity happen at the Planck scale?

This paper challenges the ubiquitous claim that physics breaks down at the Planck scale, arguing that the evidence is weaker than commonly asserted. It systematically evaluates five key arguments for this idea—including nonrenormalisability and quantum black holes—and finds all of them unconvincing. The author concludes that even the strongest argument rests on an unwarranted assumption, meaning current theories do not actually predict the necessity of quantum gravity at the Planck scale.

ArXiv

Whisper leak: a side-channel attack on large language models

A new side-channel attack, Whisper Leak, infers user prompt topics from encrypted LLM traffic by analyzing packet size and timing metadata in streaming responses. The technique is highly effective across dozens of major LLMs, achieving near-perfect topic classification even with extreme class imbalance. While mitigations like padding and batching reduce the attack's effectiveness, they do not provide complete protection, revealing a significant, industry-wide privacy vulnerability in how LLMs handle streaming data.

ArXiv

The Value of Personalized Recommendations

Researchers developed a discrete choice model using Netflix viewership data to isolate the value of personalized recommendations. Their findings show the current system increases engagement by 4% over matrix factorization and 12% over a popularity-based algorithm, while also improving consumption diversity. The study concludes that most of this lift comes from effective targeting rather than simple exposure, with the largest gains observed for mid-popularity content.

ArXiv

Show HN: DBOS Java – Postgres-Backed Durable Workflows

DBOS Transact is a lightweight, open-source library that provides durable workflows by checkpointing state in Postgres. This enables the creation of fault-tolerant, long-running applications like AI agents that can survive crashes and resume execution without losing state. Unlike externally orchestrated systems like Temporal, DBOS integrates directly into an application with no external dependencies besides Postgres, and includes features for durable queues, scheduling, and notifications.

Repo

AI-First: Practical Guidelines for Making Websites Readable by AI

first.ai is a guide for building AI-readable websites, adapting to the shift from traditional search to LLM-based discovery. It advocates for prioritizing semantic HTML, structured metadata like JSON-LD, and minimalism to make content more transparent to AI models. The project provides practical patterns to ensure web content can be easily understood, structured, and cited as a trusted source by AI.

Repo

Run Any LLM from Claude Code (GPT-5.1, Gemini, Grok,)

This Claude Code agent enables calling external LLMs like Kimi, Grok, and Gemini by mentioning the model's name within a prompt. It functions by detecting the model trigger, extracting the query, and calling the specified model via the OpenRouter API, saving the output to a local file. The underlying openrouter.sh script can also be used directly as a CLI tool, supporting context files, system prompts, and other parameters.

Repo

Show HN: I built Solveig, it turns any LLM into an assistant in your terminal

Solveig is a terminal-based AI assistant that enables safe agentic behavior from any LLM. It emphasizes safety through granular consent controls and pattern-based permissions, prioritizing file operations over direct shell execution. The tool is provider-independent, connecting to any OpenAI-compatible API (including local models), and is extensible via a Python plugin architecture.

Repo

Rethinking Graph Neural Networks for Anomaly Detection

An error occurred as the system was unable to retrieve the README file.

Repo