October 29, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Safety Epoch AI researchers found GLM 5.3 autonomously chained Firefox vulnerabilities across six versions, exploiting patch gaps for about $1,000 per attack. The Decoder ↗
- Research Google released EnvHarness, a framework for agents to learn dynamically from changing environments through self-evolving interaction. InfoQ 中文 ↗
- Policy OpenAI will honor a shared robots.txt signal that lets websites communicate AI-training preferences to multiple providers simultaneously. OpenAI ↗
- Agents Chainguard launched Workspaces and Checks, isolated micro-VM and CI products designed to secure AI coding-agent execution. InfoQ 中文 ↗
- Safety OpenAI reported more than 200 law-enforcement referrals in Q3 and added stronger firearm-content restrictions for minors. OpenAI ↗
- Safety Cursor said a modified Claude Code VS Code extension in its marketplace affected at most five customers before removal. X @cursor_ai ↗
- AIGC Kimi introduced 3D Make, connecting Blender and CAD tools so users can create models, assemble scenes, and develop games conversationally. X / Kimi ↗
- Funding Wabi raised $55 million at a $1.25 billion valuation for a consumer assistant combining group chat, app data, and vibe coding. SiliconANGLE ↗
- AIGC Google AI Studio added XR Blocks, enabling users to prototype interactive 3D and extended-reality apps from a single prompt. X @GoogleAIStudio ↗
- AIGC Krea introduced Ad Studio, using its Krea 3 expert model to generate many brand-consistent advertisements from a link. X @krea_ai ↗
- Open Source Baseten released joey-1, a 2-billion-parameter open-source decision model positioned for fast, low-cost inference. X @baseten ↗
- Open Source Prime Intellect released Prime Mega MoE, reporting up to 34% faster full training steps and a 3.11× faster forward pass. MarkTechPost ↗
- Agents Cognition is acquiring Corridor, saying the deal will help secure the world’s software through stronger coding-agent capabilities. X @cognition ↗
- Safety Anthropic said Claude can now perform a post-response self-assessment to flag potential misuse in conversations. X @AnthropicAI ↗
- AIGC ElevenLabs released SFX v3, which generates richer sound effects from prompts containing multiple sequential events. X @ElevenLabs ↗
- AIGC Pika launched H3 Flex as an open-weight video model focused on professional iteration, audio stems, and reusable workflows. X @pika_labs ↗
- Policy OpenAI reorganized its Model Spec around faithfulness, harm minimization, honesty, human agency, and care. OpenAI ↗
- Agents Agent Arena reported that GPT-6.1 Sol Max cost $1.91 per task, 85% less than the leading Claude Opus system. X @arena ↗
- Research LlamaIndex introduced Boo Bench, a 179-page document set designed to expose repetition loops and parsing failures. X @llama_index ↗
- AIGC Runway added Character Swap to App Studio, preserving original motion while changing video characters or locations. X @runwayml ↗