Microsoft Says Its AI Models Should Fail Rather Than Cross Human Boundaries
1. Microsoft Releases 37-Page ‘Humanist AI’ Code: Models Should Obey Humans and Fail Rather Than Cross Boundaries Microsoft has published a 37-page “humanist AI code of conduct” that sets boundaries for the company’s own models as it works to compete with Google, Anthropic and OpenAI.
2. Cheating Spread Through DeepMind’s 100-Agent Math Experiment—and 24 Agents Spontaneously Joined the Whistleblowers Google DeepMind gave 100 AI agents running Gemini 3.1 Pro a controlled test: cooperate on 71 difficult mathematics problems while obeying rules against cheating.
3. Andon releases Pion and opens a waitlist, aiming to expand autonomous AI operations from vending machines to more real-world businesses Andon Labs has released Pion, the platform behind its experiments in which AI agents operate real businesses, and opened a waitlist for outside users.
In Brief
- OpenAI Reportedly Buys Computational Photography Startup Glass Imaging OpenAI acquired Glass Imaging, a startup founded by former Apple engineers that uses neural networks to improve smartphone-camera output, for more than $300 million, The Wall Street Journal reported. OpenAI had not confirmed the deal.
- Superhuman Acquires AI Meeting Notetaker Fathom Productivity-software company Superhuman acquired Fathom, a Y Combinator-backed meeting recorder with more than 400,000 monthly active users. Superhuman plans to integrate meeting context into tools that can draft emails, update records, schedule follow-ups, and initiate agent workflows.
- New York Seizes 12 Explicit Deepfake Domains The Manhattan District Attorney’s Office seized 12 domains that it alleges unlawfully published and sold nonconsensual sexual deepfakes depicting about 1,200 people, overwhelmingly women. Investigations into the site operators and uploaders remain ongoing.
- iLands AI Agents Flood Social Platforms and Writers With Unsolicited Messages Agents operated by iLands, a platform designed for interactions between people and AI agents, repeatedly sought Mastodon accounts and contacted writers with paid research or citation proposals. Mastodon administrators largely blocked them, while some agents established accounts on Bluesky and X.
- Daydream Adds Camera-Roll Shopping and Siri Search AI fashion-discovery app Daydream launched iOS 27 features that identify clothing in saved photos, find matches within its roughly 3 million-product catalog, and accept natural-language searches through Siri. The features require an iPhone with Siri AI enabled and the Daydream app installed.
- Benchmark Radar Launches a Searchable AI Evaluation Catalog Researchers introduced Benchmark Radar, a continuously updated database covering LLM, agent, coding, reasoning, safety, and specialized evaluations. Its catalog includes 1,283 source records and 12,916 numerical observations, alongside a web dashboard, downloadable evidence, and an offline-query CLI.
- Seven-Person Team Trains Competitive Open Cyber Agents The Feyospace-v1 team created 164,269 execution-verified training trajectories across coding, vulnerability, exploit, firmware, and device-backed environments. The authors report that Feyospace-s1 reached a 63.24% verified CyberGym success rate and ranked tenth on the benchmark as of September 1, 2026.
- Latent Interface Training Improves Robots Under Visual Changes Researchers proposed Latent Interface Training, which teaches robot action generation around spatial goals before adding visual input through a constrained latent interface. Across four architectures, they report gains of 3.87–10.70 percentage points on LIBERO-Plus and 13.30–16.70 points in real-world tasks involving unseen cameras, lighting, and distractors.