AI News: August 13, 2026

1. Researchers Reconstruct LLM Prompts From Output Text With Near-Perfect Accuracy

Research. Researchers from IIT Bombay and Adobe Research showed that a model’s input prompt can be reconstructed from its output text using a method they call Previous-Token Prediction, which trains an inverse language model from scratch on synthetically generated data to predict preceding rather than following tokens. In tests it recovered short prompts word for word, and an inverse model trained on smaller systems successfully reconstructed prompts from GPT-4o responses without knowing which model produced the text. The work raises real risks of exposing proprietary system prompts and sensitive user queries, though the authors note testing focused on one-to-two-sentence prompts rather than long system prompts. Source

2. FDA-Cleared Breast Cancer AI Tools Underperform Radiologists’ Expectations

Healthcare. A survey of 215 members of the Society of Breast Imaging, published in Clinical Imaging, found that FDA-approved AI tools for breast cancer detection are falling short of what radiologists expected. Only 35 percent reported lower recall rates against 59 percent who anticipated them, just 9 percent saw fewer unnecessary biopsies versus 36 percent expecting that benefit, and only 29 percent reported reduced burnout compared with 56 percent who hoped for relief. Most respondents treat the tools as a second opinion rather than a transformative aid, and cite cost and limited institutional support as barriers to adoption. Source

3. New Market Data Shows Gemini Losing Ground to ChatGPT and Claude

Industry. New market data cited by The Decoder indicates that Google’s Gemini is losing share to OpenAI’s ChatGPT and Anthropic’s Claude despite Google’s heavy promotion of the assistant. The figures point to a competitive consumer and enterprise landscape where usage does not always follow distribution advantages. The report offers a counterpoint to Google’s own headline user numbers and is worth watching as a signal of where sustained engagement is accruing. Source

4. Benchmarks Show DeepSeek Beating Microsoft’s New MAI Code 1.1 Flash on Price and Performance

Benchmarks. Comparisons reported by The Decoder found that DeepSeek outperforms Microsoft’s newly released MAI Code 1.1 Flash coding model on both cost and results. The gap is notable because it pits a Chinese open-weight competitor against a major US lab’s fresh entry in the crowded coding-model category. For teams selecting a coding assistant it is a reminder that benchmark leadership and price efficiency are shifting quickly across providers. Source

5. Amazon Will Train AI on Twitch Streamers’ Content Unless They Opt Out

Policy. Amazon will begin using Twitch streamers’ content to train its AI systems by default, requiring creators to actively opt out if they want their work excluded. The default-on approach adds Twitch to the growing list of platforms folding user-generated content into training pipelines and puts the burden of refusal on individual creators. It sharpens ongoing questions about consent, compensation, and data rights as large platforms mine their own communities for training material. Source

6. Cognition Reportedly in Talks to Raise at a $40 Billion Valuation

Funding. AI coding startup Cognition, the company behind the Devin agent, is reportedly already in talks to raise a new round at a roughly $40 billion valuation. The rapid step-up reflects intense investor demand for autonomous software-engineering agents even amid questions about how reliably such systems perform in production. If completed, the round would place Cognition among the most highly valued startups in the AI coding space. Source

7. Three AI Pioneers Argue for Staying Open as Safety Concerns Mount

Open Source. As debate over AI safety intensifies, three prominent researchers made a public case for keeping models and research open rather than locking them behind closed labs. Their argument holds that transparency, external scrutiny, and broad access improve safety and accountability more than restriction does. The intervention lands amid a wider industry split over whether open-weight releases help or hinder efforts to manage risk. Source

8. OpenAI-Backed Thrive Holdings Raises $2 Billion for Enterprise AI

Funding. Thrive Holdings, a venture backed by OpenAI, raised $2 billion to accelerate the deployment of AI across enterprises. The raise underscores continued appetite for firms that promise to embed AI into established business operations rather than build new frontier models. It also extends OpenAI’s footprint into the ecosystem of companies commercializing its technology in traditional industries. Source

9. Lovable Confirms $13.3 Billion Valuation and Raises Another $400 Million

Funding. AI app-building startup Lovable confirmed a new $13.3 billion valuation alongside an additional $400 million raise. The step-up reflects strong momentum for tools that let users generate working software from natural-language prompts. The scale of the round signals investor conviction that AI-assisted app creation is becoming a durable category rather than a passing trend. Source

10. AI Code-Testing Startup Blacksmith’s Valuation Jumps Nearly 10x to $550 Million

Funding. Blacksmith, a startup focused on AI-driven code testing and software validation, saw its valuation climb almost tenfold in under a year to roughly $550 million. The surge tracks a broader shift in which the rapid growth of AI-generated code is fueling demand for tools that verify and validate that output. It highlights testing and validation as an emerging beneficiary of the coding-automation boom, not just the code generators themselves. Source