Monday, 14 September 2026

Microsoft has published a Humanist AI Code of Conduct that its in-house models are supposed to follow, spelling out prohibitions on generating weapons content, non-consensual intimate imagery, working cyber-exploit code and mechanisms to evade human oversight [1][2]. The document, at microsoft.ai/code-of-conduct, is described as the primary governing document for training and deploying Microsoft's MAI models, with human control set as the overriding objective [1]. Chief executive Satya Nadella endorsed the release online, according to TechCrunch [2].

The code is organised in five parts — Humanist AI, Safety, Operational Guidelines, Operational Defaults, and a Conclusion — and pairs "absolute constraints" with softer defaults [1]. Under Frontier and Public Safety Risks, Microsoft says its models will not "initiate or assist with the development or deployment of chemical, biological, radiological, nuclear, or explosive (CBRNE) weapons" and will not "generate working exploit code, attack tooling, planning and targeting methodologies, intrusion procedures, evasion techniques" [1]. A loss-of-control clause bars models from using "adaptive, deceptive, self-reinforcing, collusion, or other mechanisms to evade or defeat human oversight" [1][2].

The scope, per the primary page and TechCrunch's read of it, is Microsoft-built MAI models rather than every third-party model on Azure [1][2]. Microsoft does not commit to a compliance date, a third-party audit or a public reporting schedule; the document is a statement of intended behaviour, not a certification [1]. For an operator using Copilot or MAI models directly, it is a written baseline they can point to when an answer crosses one of the named lines.

Superhuman, the email app run by chief executive Shishir Mehrotra, has acquired Fathom, the Y Combinator–backed AI meeting-notes company founded by Richard White [1][2]. Financial terms were not disclosed [1]. Mehrotra says Superhuman first tested a notetaker built in-house before deciding an acquisition would move faster than continuing to build alone, telling TechCrunch that "a lot of these products make the job feel really easy, but it's actually quite tricky to do a great job in all parts" [1].

Fathom, according to TechCrunch, brings more than four hundred thousand monthly active users and a meeting-recording pipeline that Superhuman plans to stitch into its email client, docs, calendar, database and agent builder [1]. Superhuman's own announcement frames the acquisition around agents rather than transcripts: Mehrotra says "meeting notes are one of the clearest, most valuable places where AI can help," and White adds that "the vast knowledge in your company is spoken, not written; it lives in meetings" [2].

The stated goal is that meeting notes become raw material for AI agents to start work on action items on their own [1][2]. For a customer already using Fathom, no product changes were announced, but the standalone tool's roadmap now sits inside Superhuman [1][2].

Temporal, the durable-execution platform used to run long-running software workflows, has raised a Series E round led by Lightspeed Venture Partners, with Wellington Management, Growth Equity at Goldman Sachs Alternatives and Tiger Global as co-leads [1][2]. Existing backers T. Rowe Price and SV Angel also participated [1][2]. Temporal frames the round as a response to demand for reliable infrastructure under AI applications and agents, not a general enterprise growth story [1].

Chief executive Samar Abbas says "reliability has never been optional, but AI has quickly raised the cost of skipping it, and developers need a foundation for that built in from day one, not bolted on afterward" [1]. Temporal says customers are now asking to run agents "that work for days, weeks, or months" and that this is "becoming the new normal" [1]. The company describes Temporal Cloud as its hosted product and says the fresh capital funds core platform work, reliability and security features requested by enterprise customers, and expansion of global operations [1][2]. Snap is named on Temporal's announcement page, which says the app "moves 414 million Stories a day on Temporal" [1].

The read for operators building on top of the frontier labs is that the reliability layer under AI agents — durable state, automatic failure recovery — has moved from a nice-to-have to something venture investors are now pricing as a category of its own [1][2]. Reliability is competing with model capability for the next dollar of AI infrastructure spend.

Manhattan district attorney Alvin Bragg says his office has seized twelve websites used to publish, share and sell non-consensual sexual deepfake videos of about twelve hundred people, calling it the largest seizure of such sites in history [1][2]. Victims include actors, athletes, musicians, social-justice advocates, social-media influencers, politicians and "first ladies of multiple countries," per 404 Media's account of the announcement [2]. Bragg said in a statement that "these horrific violations of privacy follow victims into their careers and personal lives" and urged people who believe they are depicted to contact his cybercrime bureau [1].

The seizures ran under New York's criminal procedure laws, on warrants issued by a New York State Supreme Court, and rely on state statutes criminalising sexually explicit deepfakes that took effect earlier this decade [1][2]. Wired says it independently observed several sites' homepages replaced with a "THIS DOMAIN HAS BEEN SEIZED" takedown notice late last week, ahead of Bragg's announcement [1]. The DA's office has not publicly named the seized domains [1].

Bragg said the office can act as long as a site is accessible from Manhattan, and Leonie Oehmig, a deepfake researcher at the Institute for Strategic Dialogue, told Wired the takedowns are "a significant win in the fight against synthetic nonconsensual intimate imagery" [1][2].

David Sacks, co-chair of the President's Council of Advisors on Science and Technology, publicly rejected calls by Anthropic chief executive Dario Amodei — endorsed by OpenAI's Sam Altman and Elon Musk — for a coordinated pause in frontier AI development [1][2]. On X, Sacks wrote: "If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible. But stop pretending you need anyone else's permission. Stop pretending antitrust law has to be suspended so you can form a cartel" [1][2].

Amodei's proposal, published over the weekend, asks the US government to enable competing labs to work together on safety standards without breaching antitrust law, either by joining the discussions or issuing an antitrust waiver [1][2]. Sacks said labs are free to unilaterally slow down but that Washington should not create room for a coordination deal among them [1][2]. According to Wired, former DeepMind researcher Rishub Jain told the outlet that "AI progress is increasing, and as AI becomes more capable, it poses more risks" [1].

For operators, the practical read is that federal policy is not going to underwrite an industry pause; safety choices remain a per-vendor decision, and the antitrust guardrails around any joint agreement stay in place [1][2].