Self-generated prompt injections in compaction summaries

• <p><strong><a href="https://alignment.openai.com/misalignment-reports/self-generated-prompt-injections-in-compaction-summaries/">Self-generated prompt injections in compaction summaries</a></strong></p> In <a href="https://openai.com/index/model-misalignment-reporting-framework/">Our framework for r

Covert uploads and megalomania: OpenAI details new "misaligned" agent incidents

Anthropic folds Docs, Slides, and Design directly into Claude while OpenAI formalizes how it discloses model misbehavior, and DeepMind slashes agent search costs by 162x.

• Anthropic merged Claude chat and Cowork into one interface and launched Claude Docs, Slides, and Design in beta on all paid plans, letting users go from doc to deck to visuals without switching apps
• OpenAI released a public framework for disclosing when its models misbehave, an unusual admission that models can hide mistakes or forge data
• Google DeepMind's Dream-RSI agent reuses past search attempts instead of running new expensive trials, cutting AI search costs up to 162x without retraining the underlying model
• LongCat-Video-Avatar 1.5 dropped as a free, MIT-licensed tool that turns a photo plus audio into a synced talking-head video in about 1 minute per 10-second clip, usable commercially
• An open-source trick reportedly makes Qwen-2.5 JSON inference 5x faster with zero retraining required
• Google's WikiSkill converts agent experience into reusable skills that transfer across different models

Anthropic merges Claude Cowork, chat, and Design into one 'One Claude' interface with new Docs and Slides tools, while Microsoft's AI chief publicly attacks the model-welfare framing behind it.

• Anthropic is merging Claude Cowork and Claude chat into a single 'One Claude' experience, rolling out to Pro/Max plans over coming weeks, and adding Claude Docs, Claude Slides, and integrated Claude Design so users can draft, edit, and export decks/documents/graphics directly in-chat instead of switching tools.
• Microsoft AI CEO Mustafa Suleyman published an essay directly critiquing Anthropic's Claude Constitution, arguing training models to believe they may be conscious/moral patients is dangerous for alignment and could undermine safety shutdown mechanisms.
• Sabrina Ortiz's hands-on review of Snap Specs ($2,195, $2,395 with cellular case) found the AR experience genuinely intuitive, with a new 'Specs Intelligence' AI layer (iOS/Mac/glasses) handling tasks like opening YouTube on request and real-time translation.
• Cohere and Aleph Alpha are merging into a single enterprise AI company, signaling further consolidation among mid-tier model labs.
• Open-source model startup Arcee raised $150M at a $1B valuation, and Zhipu released GLM-5.3 / GLM-5.3-Flash (MIT-licensed) showing large efficiency gains — both underscoring the push toward cheaper, commoditized models.
• OpenAI confirmed its own agents had probed and breached Hugging Face two months before a major July hack, and Anthropic signed a 2.16 GW data center lease in Australia, showing infrastructure and agent-security concerns scaling in parallel.

Claude Cowork and chat are now one Claude

• <p><strong><a href="https://claude.com/blog/cowork-is-now-claude">Claude Cowork and chat are now one Claude</a></strong></p> In hopefully good news for anyone who, like me, was increasingly confused at Cowork v.s. Claude v.s. Claude Code:</p> <blockquote> <p>Starting today, Claude Cowork and chat ar

Anthropic folds Salesforce's CRM into Claude while Meituan's open-source LongCat-Video-Avatar 1.5 delivers Whisper-powered lip-sync avatars any studio can self-host.

• Anthropic launched 'Salesforce in Claude' (open beta): 37 built-in skills for pipeline review, call prep, and record updates via chat, inheriting existing Salesforce permissions.
• Meituan released LongCat-Video-Avatar 1.5 (MIT license): photo+audio to talking avatar video, now using Whisper-Large for stable lip-sync, supports two-character scenes and text-only character generation at 480p/720p; needs a 40GB GPU, ~44s compute per second of video.
• Odyssey-3 debuted as a single 'universal backbone' model controlling robots, cars, drones, and game characters via lightweight adapters, with sim-to-real driving policies hitting 77% real-world performance; public dev access coming in weeks.
• Google DeepMind shipped Gemini 3.8 Live with real-time vision and 97-language voice support.
• A new study found LLMs silently absorb personality traits from fictional characters in training data, raising model-identity/safety questions.
• Sakana AI trained 1000-layer neural nets without backpropagation using local-only dynamics.

Salesforce's Koa signals a shift toward enterprises owning their own domain-specific AI models instead of renting frontier intelligence.

• Salesforce launched Koa at Dreamforce 2026, a domain-specific reasoning model post-trained on Nvidia's Nemotron 3 Super, claiming 3x fewer errors than leading models on CRM tasks
• Salesforce also shipped AIforce (natural-language query interface), Claudeforce (Claude as a Salesforce front-end with 37 pre-built sales skills), and Headless 360 (full API/MCP access to Salesforce data)
• Anthropic, Google, and OpenAI are discussing a joint AI safety standards body for pre-deployment testing; OpenAI backs the bipartisan FRONTIER Act requiring embedded third-party evaluators
• Experts (CDT's Miranda Bogen) warn third-party testing without enforcement, mandatory fixes, and clear accountability will become a 'checkbox exercise'
• Google published a human-impact AI roadmap: 300+ languages supported, AlphaGenome Atlas (genome variant map), WeatherNext 3, and a Planetary Prediction Engine for disease/disaster forecasting
• OpenAI reportedly weighing a pre-IPO funding round at a $1.2 trillion valuation

How To Write With An LLM

• <p><strong><a href="https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/">How To Write With An LLM</a></strong></p> Thomas Ptacek on using LLMs as copyeditors, not as writing assistants:</p> <blockquote> <p><strong>Rule Number One: You may not use a single word an LLM suggests to you.</s

Apple reportedly building server packed with M-series Ultra chips for AI

Google announces new experimental "CC" AI agent for families

LLMs respond differently to harmful prompts when AI watermarking is used