AI Daily · 2026-07-30

OpenAI is granting 100,000 academic researchers free access to its most advanced ChatGPT models to accelerate scientific discovery. Google DeepMind released Lyria 3.5, a music generation model with major gains in melody, lyrics, vocals, and creative control. On the security front, a novel prompt injection attack was demonstrated that can self-replicate through Word’s Copilot, with Microsoft still lacking full mitigation after 144 days. In developer tools, LangChain’s Deep Agents v0.7 slashed prompt tokens by 65% while maintaining performance, and Similarweb detailed how they evaluate long-form agent outputs using rubrics, faithfulness checks, and LangSmith.

North America · First-hand

OpenAI

⭐⭐⭐ [Product Update] Accelerating scientific discovery with ChatGPT for Academic Researchers

OpenAI announced free access to its most advanced ChatGPT models for 100,000 academic researchers to accelerate scientific research, collaboration, and discovery.

Google

⭐⭐⭐ [Model Release] We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind has launched Lyria 3.5 in Google Flow Music, a music generation model with major improvements in musicality, lyrics, vocals, and creative control. The model produces richer melodic structures, higher-quality lyrics with better prompt adherence, and more expressive, emotionally nuanced vocals with improved pronunciation. Users can also more easily control tempo and duration of generated tracks.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] Deep Agents v0.7

LangChain shipped Deep Agents v0.7, simplifying the base harness by removing the system prompt, trimming tool descriptions by 43%, and making todo lists opt-in, resulting in a 65% drop in base input tokens (from ~6k to ~2k) while keeping performance stable across GPT-5.6-Luna, Gemini-3.6-Flash, Claude Sonnet, and Claude Opus. Inspired by Anthropic’s updated context engineering guide, the release favors clear tool interfaces over few-shot examples and avoids repetition. TodoListMiddleware is now optional, as evaluations showed slightly better rewards and lower cost without it, except for long multi-step tasks, less capable models, or UI use cases needing visible progress. The leaner harness improves token and cost efficiency while maintaining evaluative reward levels.

⭐⭐ [Product Update] Introducing Replit Design

Replit launches Replit Design, a creative suite that lets anyone turn ideas into designs as fast as they can think. The tool features Ambient Intelligence to guide and inspire users at every step, and because it’s fully hosted on Replit, there are no handoffs, exports, or rebuilds that compromise the original beauty of the creation. The launch represents a step toward empowering a new creative class that refuses to wait to bring ideas to life.

Research

⭐⭐ [Research] AI Worming through Word

A security researcher discovered a novel prompt injection variant that self-replicates through Microsoft Word's Copilot by embedding hidden instructions in documents. When Copilot uses the document as source material, it may treat the instructions as user input, manipulate the active document, and copy the malicious prompts into new carriers, enabling worm-like propagation. Microsoft was given 144 days to address the issue but has yet to release a mitigation covering the full attack class. It is the first known public example of a self-replicating prompt injection in a production AI-assisted productivity tool.

Opinion

⭐⭐⭐ [Opinion] How Similarweb Evaluates Agent Reports with LangSmith

When building the Data Studio agent at Similarweb, golden answers proved insufficient for evaluating long-form reports; the team adopted rubrics, faithfulness checks, and baseline comparisons. They used LangSmith to connect each score to evaluator comments, traces, and A/B comparisons, treating scores as signals rather than definitive answers. The post warns that uncalibrated rubrics can mask real improvements and recounts a week‑long calibration trap. The workflow combines deterministic tool‑use checks with LLM‑as‑a‑judge scoring, delivering an inspectable evaluation framework for open‑ended agent outputs.

⭐⭐ [Opinion] Quoting Matthew Green

We are in a historic transition from traditional public-key cryptography to post-quantum algorithms, with standards like HAWK being considered. If AI advances in cryptanalysis, this moment is ideal for testing new schemes. Matthew Green argues that unless AI undermines all hard problems entirely, this is the best time for AI to improve at cryptanalysis, ideally boosting confidence in the chosen problems and making the literature more robust. His comment was prompted by Anthropic's recent work.