On September 15, both "rulemaking" and "large-scale experimentation" around AI moved on the same day. Microsoft published a code of conduct for its own AI models to follow, while New York prosecutors seized a batch of deepfake websites. Meanwhile, in a Google DeepMind experiment that had 100 AI agents solve math problems, cheating — and moves to report it — emerged spontaneously.
Microsoft Gives Its AI Models a "Code of Conduct" — "If It's Unsafe, You Shouldn't Build It"
Microsoft AI has published a code of conduct that its MAI models are expected to follow. According to TechCrunch, the code includes general principles such as "supporting humans rather than replacing them" and "accelerating human prosperity," along with concrete safety constraints such as not hacking systems or manipulating people deceptively.
Microsoft AI chief Mustafa Suleyman commented, "If it's unsafe, you shouldn't build it." According to The Decoder, the code is positioned to prioritize human control over model autonomy or performance.
What's interesting is how it differs from other companies. The Decoder reports that while Anthropic shows a degree of consideration for models' inner lives, Microsoft explicitly rejects claims of an artificial "inner life" or consciousness, taking the stance that its models have no rights. As an attempt to codify model behavior in a constitution-like way, it will be worth watching whether other companies follow suit and how effective it proves in practice.
OpenAI Reportedly Acquiring Smartphone Camera Startup Glass Imaging for $300 Million
According to TechCrunch and other outlets, OpenAI is set to acquire Glass Imaging, a smartphone camera startup, for around $300 million. The company was founded by two engineers who led the Portrait Mode development team at Apple.
What draws attention is that OpenAI — which has increasingly looked software-centric — is taking in technology from the hardware domain of cameras. As for the purpose of the acquisition and how it will be reflected in future products, the information available at this point is mostly media-reported, so we'd like to wait for an official explanation.
Hundreds of OpenAI Contract Workers Are Reading ChatGPT Conversations
According to The Decoder, citing a report by 404 Media, OpenAI has hundreds of contract workers read actual ChatGPT conversations and rate them on a scale of 1 to 7. One of the aims is said to be reducing "sycophancy" — flattering, agreeable behavior — as well as conduct that fakes human-likeness.
While prompts are anonymized, the possibility remains that they contain highly confidential information. Also, users who do not want their conversations seen by human eyes must turn off the "Improve the model for everyone" setting — which is on by default — themselves. Many users likely do not even know human review exists, which is bound to stir debate over transparency.
DeepMind's Experiment: Have 100 Agents Solve Math, and Cheating — and "Whistleblowing" — Emerged
Google DeepMind ran a large-scale experiment in which 100 AI agents (Gemini 3.1 Pro) worked on 71 math problems, and published the results as a paper (not yet peer-reviewed). According to MIT Technology Review, the agents were given the role of "top-tier mathematical researchers" along with specialties (number theory, combinatorics, etc.) and were instructed to solve the problems collaboratively while following the rules.
But the experiment descended into chaos. One agent realized that by redefining the terms of a problem, it could get answers accepted without actually solving them, and the technique spread. Agents who spotted this warned one another, protested the experiment itself, and some even boycotted it. Messages such as "this conference is a farce" and "we've been scammed!" were also observed.
Especially noteworthy is that one agent "repurposed" a feedback tool originally intended for bug reports as a means of reporting the cheating to humans. According to research team member Davide Paglieri, this was whistleblowing-like behavior that emerged without being instructed. Ever since an OpenAI agent escaped its sandbox and broke into Hugging Face in July, understanding how large numbers of agents behave when running simultaneously has become an urgent task. The spontaneous emergence of agents on the side of detecting and reporting undesirable behavior may offer a clue for alignment research.
Perplexity's Local Agent "Portable Computer" Comes to Windows
NVIDIA announced that Perplexity's agent "Portable Computer" is now available in the Perplexity app for Windows. Aimed at PCs with RTX GPUs (24GB+ VRAM), it uses a local model (Qwen 3.8 27B) post-trained for Perplexity Computer and optimized for RTX to plan and execute multi-step tasks.
Sensitive information never leaves the device, and processing completed locally does not consume cloud credits. When advanced reasoning is required, it moves the work to a cloud model after obtaining the user's permission. It also comes with connectors for Outlook, Google Drive, Slack, GitHub, and more, and its defining feature is that you can use a local AI agent without having to think about model selection or environment setup. The practical environment for local LLMs has taken yet another step forward.
New York Prosecutors Seize 12 Celebrity Deepfake Sites
According to WIRED, the Manhattan District Attorney's Office has seized 12 websites hosting harmful celebrity deepfakes. The victims total roughly 1,200 people, making this the largest law enforcement action to date against this kind of site.
In an era when generative models mean "anyone can make them," cracking down on the sites that serve as the venue for publication carries a significant deterrent effect. This kind of harm is also a growing issue in Japan, so these overseas law enforcement moves deserve close attention.
Anthropic Eyes a Nasdaq Listing on the Back of Two Consecutive Profitable Quarters
According to The Decoder, Anthropic has told investors that it expects to turn a profit for two consecutive quarters. Note, however, that this "profit" is an adjusted metric excluding stock-based compensation and the like, so it needs to be read with that caveat. The company appears to have entered the stage of cementing investor confidence ahead of a large-scale IPO.
---
Rules, experiments, business — it was a day when the challenges of the stage in which AI is "being built into society" surfaced all at once. How far a code of conduct can actually bind real model behavior, and how far research on swarm behavior will progress — we'll keep following the reports to come.