the-decoder.com

OpenAI trained an internal AI model called GPT-Red to automatically find security flaws in GPT models. GPT-Red simulates prompt injections and other attacks where malicious instructions hide in emails, websites, or files. Trained via self-play reinforcement learning, GPT-Red attacks while defender models block, and both improve over time. It finds successful attacks in 84 percent of test scenarios versus 13 percent for human red teamers.

Nano Banana Pro prompted by THE DECODER A researcher at the University of Pennsylvania has used OpenAI's language model GPT-5.6 Sol Pro to solve a long-standing open problem in statistics. The breakthrough involved disproving the assumption that the widely used Benjamini-Hochberg method for controlling false positives always works reliably, even when applied to continuous data.

PrismML argues that modern AI apps increasingly need powerful models to run locally. An agent may make hundreds of model calls in sequence, each carrying context, producing structured output, and feeding into the next step. In the cloud, per-token costs pile up, every call adds network latency, and intermediate results, tool calls, and private data such as screen content or documents all leave the device.

Spotify is expanding its AI voice interface. Premium subscribers can now talk to or text the service directly inside the app. "Play some artists I haven't heard before," "make it more upbeat," or "just his recent stuff" are some of the example commands from the announcement. In the playback view, users can ask questions about songs, artists, podcasts, and audiobooks.

Nano Banana Pro prompted by THE DECODER OpenAI's coding tool Codex encrypts the internal instructions exchanged between AI agents. Developers now see only unreadable strings and can no longer track how tasks get delegated. Encryption is mandatory for the larger GPT-5.6 models Sol and Terra, but users report it sometimes causes failed handoffs. OpenAI hasn't explained the change.

Former and current Meta employees are suing the company over allegedly discriminatory layoffs carried out by AI. According to a lawsuit filed in a federal court in California, Meta allegedly used internal AI systems to generate layoff lists when it cut 8,000 employees in May. The systems supposedly targeted employees with disabilities and those on protected medical, family, or parental leave at disproportionate rates.

OpenAI is developing a screenless smart speaker designed as a personalized AI companion for the home. The device combines a camera and sensors with OpenAI's new GPT-Live voice mode. Moving mechanical parts are meant to make the speaker feel alive, while proactive features help it anticipate and respond to users' needs. Ask about this article… OpenAI is planning a portable, screenless smart speaker as an AI companion for the home.

Anthropic is rolling out Claude for Teachers, a free offering for verified K-12 educators at US schools. It includes access to Claude along with Claude Code and the Cowork agent feature, a library of teaching-specific skills, and connections to curriculum-aligned content. The company says it covers education standards across all 50 states. Teachers can use it to plan lessons, differentiate materials, or analyze student data.

Nano Banana Pro prompted by THE DECODER DeepSeek is planning a new funding round at a $71 billion valuation to build its own data centers and buy AI chips. The fresh capital would fund an aggressive expansion strategy. Permanent rock-bottom pricing on its new V4 models is driving rapid growth among US businesses. While DeepSeek still trails Western frontier models on performance, domestic pressure is mounting too, with Chinese rivals like Zhipu AI, MiniMax, and Moonshot AI all pushing hard.

Google is adding AI image generation directly into Search's AI Overviews. When no matching image exists on the web, users can generate one by typing a text prompt into the search bar. The feature runs on Google's new "Nano Banana 2 Lite" image model, which prioritizes speed and cost over quality. The rollout starts in the coming weeks in English across all regions that already support . For the open web, this likely means fewer clicks.