September 28, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Open Source Google Research open-sourced RRSI, an agent framework that improves its own harness while explicitly limiting overfitting. MarkTechPost ↗
- Research IQuest-Q1 reportedly identifies reinforcement-learning data bugs and generates playable mini-games directly from prompts. 量子位 ↗
- Agents Genspark added Claude Sonnet 5.5 to its chat, code-agent, and Claw products, claiming faster output and lower task costs. X @genspark_ai ↗
- Funding Noin Intelligence completed its fifth financing round within a year, raising several hundred million yuan and over one billion yuan cumulatively. 量子位 ↗
- Safety OpenAI published its approach to securing frontier reinforcement-learning training runs against theft, misuse, and operational compromise. X @OpenAI ↗
- Open Source Qdrant introduced Constella, a research preview allowing query-embedding model changes without re-embedding stored documents. X @qdrant_engine ↗
- Models Alibaba released Qwen-Audio-3.1-Realtime, a full-duplex voice model trained to reason, act, and decide when to speak. MarkTechPost ↗
- Research A new arXiv paper proposed SkillCascade-Bench to test harmful interactions across multiple skills in agent systems. arXiv ↗
- Safety Omdia said AI-agent identity security requires layered defenses rather than a single control point. SiliconANGLE ↗
- Agents GitHub showed how a marketer turned repetitive event operations into an automated workflow triggered by a single issue form. X @github ↗
- Agents Okta outlined a shared architecture for securing agent runtimes as autonomous systems gain access to enterprise tools and identities. SiliconANGLE ↗
- Open Source NVIDIA spotlighted Nemotron Labs, a small Korean team that reached a top-five score in an open-model evaluation. X @NVIDIAAI ↗
- AIGC Krea added Seedance 2.5 Draft Mode, offering 480p generations before users switch selected scenes to 1080p. X @krea_ai ↗
- AIGC Pika highlighted Secret Level projects built around generative media, including animated Coca-Cola and original narrative productions. X @pika_labs ↗
- AIGC Kuaishou says Kling 4.0 will launch in October with 4K HDR and 30-second video generation. IT之家 ↗
- Agents Fireworks AI introduced FireRouter, routing routine requests to open models while reserving Claude Opus for harder tasks. X / Fireworks AI ↗
- AIGC Runway added ElevenLabs v4, offering more expressive narration and character dialogue inside its video-creation platform. X / Runway ↗
- Agents Cloudflare introduced an agent-development lifecycle intended to replace traditional software-development lifecycles with continuous agent evaluation and governance. InfoQ 中文 ↗
- Models Grok 4.7 became available through Amazon Bedrock, extending xAI’s model reach into AWS enterprise infrastructure. AWS Machine Learning ↗
- Industry Character.AI began rolling out ShortSqueak and PipSqueak 3 as new free Chat Styles for conversational customization. X / Character.AI ↗
- Funding Modal Labs is reportedly nearing a $750 million funding round at a $15.75 billion valuation. TechCrunch ↗
- Agents Replit promoted Ask Replit, which turns natural-language questions into charts, data pipelines, or working automations. X / Replit ↗
- Safety 1Password introduced task-based verification to limit AI-agent access according to the specific action being performed. SiliconANGLE ↗
- Agents Shopify opened checkout flows to browser-based AI agents, allowing automated systems to complete purchases on merchants’ behalf. TechCrunch ↗
- Research LlamaIndex said document parsing still challenges frontier vision-language models because forms encode fields, groups, and spatial relationships. X / LlamaIndex ↗
- Agents Google is replacing Gemini’s Gems with reusable skills, shifting customization toward more structured, action-oriented workflows. TechCrunch ↗
- Research Goodfire argued that sparse autoencoders are best for exploratory interpretability, while targeted probes work better for known questions. X / Goodfire AI ↗
- Safety Perplexity said eight of ten third-party sandbox platforms it tested had similar network-policy bypass vulnerabilities. X / Perplexity ↗
- AIGC ElevenLabs announced v4 and v4 Turbo improvements, including IPA pronunciation control, expressive delivery tags, and discounted API access. X / ElevenLabs ↗