September 15th, 2026: Model codes, advisor agents, and deepfake enforcement
AI controls moved into model rules, regulated workflows, and domain seizures, while access and oversight became practical design questions.
Control stopped being a slogan in several of today’s strongest stories. Microsoft drafted behavioral rules, Anthropic put approval gates around financial-advisor workflows, and Manhattan prosecutors seized distribution infrastructure. Each acts at a different enforcement point—model behavior, workflow, or distribution—while the smaller developments expose the same tension in vehicle engineering, usage quotas, training-data bargains, and still-private standards talks.
Lead developments
1. Microsoft turns “humanist AI” into a draft rulebook
Microsoft AI published a draft Code of Conduct for the MAI models it develops and opened a six-week public consultation. The text translates “people matter more than AI” into proposed requirements: models should act only within scoped authorization, remain interruptible and traceable, avoid self-assigned goals, and never resist shutdown. It also sets absolute prohibitions covering areas such as weapons, child exploitation, nonconsensual sexual content, and mass manipulation. The important boundary is timing: Microsoft says the draft does not govern training today and is intended to guide development from 2027 after revision. For model buyers, a written rulebook creates a testable target, but it is not yet evidence that a deployed system meets it.
Watch: The final text, published evaluation methods, and whether the rules cover products that combine MAI with third-party models.
Sources: Microsoft AI’s announcement and consultation, the full draft Code of Conduct, Axios’s interview on the competitive trade-offs
2. Claude enters the financial-advisor workflow with approval gates
Anthropic launched Claude for Financial Advisors, a Cowork plugin that connects to custodians, portfolio systems, customer databases, planning tools, and meeting records. Its practical contribution is orchestration: Claude can gather account context, prepare meeting briefs, screen client-facing language against the SEC Marketing Rule, and stage updates or drafts for review. Anthropic says investment recommendations, client communications, compliance decisions, and other regulated work remain subject to human approval; it recommends Enterprise because its audit logs support recordkeeping. That design moves the competition beyond answer quality toward permissions, provenance, and review state across a workflow. It also leaves responsibility with the adviser and firm: a compliance screen can surface an issue, but it does not make the underlying judgment.
Watch: Connector permissions, audit-log detail, error rates on real portfolios, and measured time saved after mandatory review.
Sources: Anthropic’s launch and workflow description, Bloomberg’s report on the financial-industry expansion
3. Manhattan prosecutors seize 12 deepfake-porn domains
The Manhattan District Attorney’s Office seized 12 domains that it says distributed and sold nonconsensual sexual deepfakes depicting about 1,200 people, overwhelmingly women. The action was carried out under court-authorized seizure warrants; investigations into site operators and uploaders remain open, and the office announced no arrests. WIRED independently observed several homepages displaying seizure notices before the announcement. This matters because enforcement moved upstream from asking platforms to remove individual files to disabling distribution infrastructure. A domain seizure is still one layer of a wider market: generation tools, discovery channels, hosting, payment systems, and affiliates can help the same material reappear. The result is a meaningful interruption, not proof that the network behind the sites has been dismantled.
Watch: Operator charges, attempts to rehost the collections, and coordinated action against hosting and payment infrastructure.
Sources: the Manhattan District Attorney’s seizure announcement, WIRED’s reporting and independent site observations
More signals
4. Pony.ai brings a Gen-4 robotruck to Europe’s freight market
Pony.ai and GAC debuted a battery-electric Level 4 truck with redundant steering, braking, power, compute, and sensing systems. Pony.ai schedules volume production for later in 2026 and is discussing European road tests. Those talks are not permits or orders, and its cost and energy-saving figures remain vendor projections; the near-term test is regulatory access under defined routes and conditions.
Sources: Pony.ai’s Gen-4 robotruck announcement, Reuters on the vehicle and European test discussions
5. Claude Code keeps part of a promotion, while weekly capacity falls from its peak
Anthropic ended Claude Code’s 50% promotional weekly boost and made a 25% increase over the old baseline permanent for eligible paid plans. That new level is about 16.7% below the temporary peak; five-hour limits are unchanged, and Free and consumption-based Enterprise plans are excluded. Teams budgeting unattended coding work should plan against the weekly meter, not the promotional experience.
Sources: Anthropic’s weekly-limit update and eligibility rules, the Claude Code community’s launch-day discussion
6. Bolt Forge trades cheaper open-model building for opt-in training data
Bolt added Forge, a research-preview agent using open-weight models led by GLM 5.3 Flash. Individual Pro users can receive up to 50 times more Forge usage through October 14 by opting to share sessions that Bolt says are anonymized and stripped of secrets for training with Arcee AI. The benchmark is Bolt’s own, and Teams and Enterprise are excluded, so this is best treated as an experimentation tier.
Sources: Bolt’s Forge launch, benchmark, and data-use terms, an independent changelog summary of the preview
7. Reports say three frontier labs are exploring shared standards
The Information and CNN report that Anthropic, OpenAI, and Google have discussed a shared body for AI testing and auditing, with a working group meeting as recently as the past week. No organization, charter, membership list, timetable, or binding rule has been announced, and the companies did not confirm the talks to CNN. A public proposal must show both enforcement power and room for smaller labs.
Sources: The Information’s original report on the working group, CNN’s report on the discussions
What to watch next
The next turn belongs to artifacts outsiders can inspect: published model evaluations, permission and audit trails for regulated agents, durable disruption of abusive networks, and third-party results for cost and safety claims. A public standards-body charter would be especially revealing if it names who sets tests, who can join, and what happens when a member fails.
Sources
- Microsoft AI announcement and consultation
- Microsoft AI draft Code of Conduct
- Axios interview on the Microsoft AI code
- Claude for Financial Advisors announcement
- Bloomberg report on the financial-advisor launch
- Manhattan District Attorney seizure announcement
- WIRED report and independent site observations
- Pony.ai Gen-4 robotruck announcement
- Reuters report on Pony.ai European test discussions
- Anthropic weekly-limit update
- Claude Code community discussion of the limit change
- Bolt Forge launch and data-use terms
- Bolt Forge independent changelog summary
- The Information report on AI standards-body talks
- CNN report on AI standards-body talks