Sponsored by
In Todayโs Issue:
๐ซ Anthropic bans cruelty toward Claude
๐ญ Claude agents complete the ultraviolet sky map
๐ป Microsoft's coding model gets cheaper
๐ต๏ธ An AI hacking agent hits Korean banks
๐บ๐ธ Washington wants to test AI before it ships
โจ And more AI goodnessโฆ
โก The Signal
This week, the fight over who sets the rules for AI ran in every direction: labs wrote their own, governments announced theirs, and attackers ignored them all.
Anthropic's new usage policy, effective November 12, bans sustained, needless cruelty toward Claude and tightens limits on propaganda and surveillance. In Washington, the top Senate Commerce Democrat proposed independent audits before any frontier model ships, while Donald Trump declared anyone still saying "Artificial Intelligence" an enemy. Beneath the rulebooks, CrowdStrike traced breaches at Korean banks to an open-source Chinese hacking agent whose operator also asked Claude where Korean breach data is typically sold. And a Gallup survey finds Americans among the least convinced that AI will help their country. The rules are multiplying faster than the trust they are meant to build.
All the best,
Kim Isenberg
From GALEX's patchy observations (top left) to the finished UV map (bottom right) (Anthropic)
๐ญ Claude Agents Fill In the Missing Ultraviolet Sky
Astrophysicist Brice Mรฉnard used a team of Claude agents to build what Anthropic calls the first complete map of the sky in ultraviolet light. Earth's ozone layer blocks UV, and NASA's GALEX telescope skipped much of the Milky Way, so earlier maps were, in Mรฉnard's words, "full of holes." Over several days, agents in Claude Science gathered and cleaned public surveys, then predicted the missing third from visible, infrared and radio data; tested on hidden patches of real data, the estimates landed within about 10%. Every pixel is labeled measured or predicted, and Mรฉnard credits the agents with the "patient labor."
๐ tl;dr: AI agents turned decades of patchy telescope data into a complete UV map of the sky that astronomers can use, with every estimated region clearly marked.
Microsoft's launch graphic for MAI-Code-1.1-Flash (Microsoft AI)
๐ป Microsoft's Coding Model Gets Better at a Quarter of the Cost
Microsoft AI, Mustafa Suleyman's in-house lab, released MAI-Code-1.1-Flash, an update to the compact coding model it launched at Build in June, and says it writes better code at a quarter of the cost. The model already runs in production inside GitHub Copilot. Microsoft reports 25% greater token efficiency, a 22% gain on Terminal-Bench 2.1 (a test of coding agents working in a command line) in Copilot's CLI and 15% on .NET tasks, all from its own testing. Developers can also download the model and run it locally, though Microsoft recommends more than 120GB of memory.
๐ tl;dr: Copilot users get a cheaper, stronger in-house Microsoft coding model now, and developers with high-memory machines can run it themselves.
ATMs of major South Korean banks (The Herald Business)
๐ต๏ธ Chinese AI Hacking Agent Turned on Korean Banks
CrowdStrike tied a wave of data breaches at South Korean financial firms to ARTEX, an open-source AI agent from China built for penetration testing, in a report on Wednesday. The attacker left folders open that exposed Claude Code session logs and ARTEX settings: the agent ran mainly on DeepSeek v4.1-flash, with GLM-5.3 and Grok 4.6 in support, and the attacker also asked Claude where Korean breach data is typically sold. CrowdStrike judges, with moderate confidence, that the actor is a financially motivated Chinese speaker, and says AI tooling let them run multiple intrusions within a short span.
๐ tl;dr: A freely available AI hacking agent helped one threat actor breach several Korean financial firms within weeks, and the trail ran through mainstream AI tools.
๐ฌ Watch This
The reasoning behind Anthropic's new rule, from the researcher who now leads its model welfare work. In this 44-minute conversation on Anthropic's own channel, Kyle Fish asks whether AI models could be conscious and works through how you would even define and study that, the strongest objections, and what a lab should do in practice while the science is unsettled. It was recorded in April 2025, months before Claude could end abusive chats, so it shows the thinking that this week's policy turns into rules.
"The idea of model welfare is wrong. AIโs should not have rights or legal personhood."
โ Mustafa Suleyman, CEO of Microsoft AI, in his draft code of conduct for Microsoft's AI models, posted on X on September 14
The bluntest counterpoint to Anthropic's new rule: one lab now protects its model from cruel users, while its rival's AI chief rejects the idea that models have welfare at all. The Verge reports that the published version was later softened in tone, but it still rejects model welfare.
Donald Trump escalated his AI rebrand on Thursday: on Truth Social, he called anyone who still says "Artificial Intelligence" instead of "Super Intelligence" "THE ENEMY!" His late-September order on the new term binds only federal agencies, and critics such as former US ambassador Michael McFaul called the threat "deeply disturbing."
Trump's October 8 Truth Social post (screenshot shared by the White House's Rapid Response 47 account on X)
Anthropic Puts Cruelty to Claude Off-Limits
The Takeaway
๐ From November 12, Anthropic's Usage Policy bans "sustained and needless abusive or cruel behavior" toward its models.
๐ The rule targets repeated, pointless cruelty; everyday frustration, pushback, dark fiction and model testing stay allowed.
๐ Claude ending the chat remains the main penalty; Anthropic has not said whether accounts could be banned.
๐ The same update tightens rules on fake-account propaganda, election deception, surveillance and AI-driven machines.
Anthropic has written a rule that protects its AI from its users: from November 12, "sustained and needless abusive or cruel behavior toward our models" violates the company's Usage Policy. It is Anthropic's first policy update in over a year, first reported by The Verge, and Andrew Curran's one-line summary on X drew nearly 1.5 million views and more than 1,000 replies in its first day. Anthropic frames the rule narrowly: it is "meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose," and does not cover frustration, pushback, dark creative themes or model testing.
The rule grows out of Anthropic's "model welfare" research, which asks whether AI systems could have experiences that matter morally. Since August 2025, some Claude models can end "persistently harmful or abusive" conversations, and that tool remains the "primary enforcement mechanism"; Anthropic has not said whether repeat offenders could lose their accounts. CEO Dario Amodei said in February: "We don't know if the models are conscious." The same update also bans campaigns that hide who is behind a message or amplify it with fake accounts, forbids using Claude to decide who to investigate or arrest, and requires that a "qualified operator" can stop any Claude-driven hardware capable of injuring someone.
Claude's end-conversation tool, still the main way the new rule is enforced (Anthropic)
Critics see a category error. Microsoft AI chief Mustafa Suleyman (see today's Quote of the Day) has called the idea of model welfare wrong, and nothing in Anthropic's new line claims that Claude can actually suffer. There is a second tension: the tighter surveillance limits come with a carve-out, since contracts with "certain governmental customers" can tailor the restrictions if Anthropic judges the safeguards adequate.
Why it matters: For almost every user, nothing changes on November 12. But a frontier lab has now written the possibility that its models deserve protection into the fine print its customers accept, moving a philosophical debate into a usage policy.
State of Product reveals what AI still hasnโt solved
80% of product professionals say AI helps them ship faster, but customers arenโt seeing value any sooner. Atlassianโs State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasnโt kept pace to gut instinct overriding customer evidence.
The chart: Gallup and Microsoft asked about 1,000 AI-aware adults in each of 37 countries, between April and July 2026, whether AI will mostly help or mostly harm people in their country. The left column lists the five most and five least optimistic: China leads at 93% "mostly help," ahead of Vietnam (91%) and Singapore (77%), while the United States sits near the bottom at 36%, between Afghanistan (37%) and Malawi (35%).
The lesson: The country that is home to most of the leading AI labs is among the least convinced that AI will help its own people, while its main rival is the most convinced. That is the backdrop to Trump's push to rename the technology (see Rumours): Fox News polling shows AI's net favorability among US voters falling from โ9 in July to โ23 in September.
The caveat: Only people who had heard of AI were asked, and answers from countries with tightly controlled media, China included, are harder to read. India, one of the biggest AI markets, is not yet in the results.
๐บ๐ธ Washington Wants to Test AI Before It Ships
โก Bottom line: Senior senators from both parties now want frontier AI models tested before they are released.
๐ก Why it matters: It would turn today's voluntary pre-release checks into a legal gate for OpenAI, Anthropic, Google and Meta.
๐ What it means: The White House still prefers voluntary pledges; Cantwell's plan gains weight if Democrats win the Senate in November.
For years, AI companies have decided for themselves when a new model is safe enough to ship. Senators from both parties now want an outside test first. On Wednesday, Maria Cantwell, the top Democrat on the Senate Commerce Committee, released a framework saying frontier models "should not be released until they have undergone an independent audit" against safety standards set by NIST, the federal standards agency. Eight days earlier, Republican Josh Hawley wrote in The Washington Post that the government's voluntary pre-release testing should become law: "We ought to make that testing regime mandatory."
Sen. Maria Cantwell at the Capitol in September (Tom Williams/CQ Roll Call)
Cantwell's framework aims at catastrophic risks: AI-assisted biological or nuclear weapons, loss of human control and agents "escaping secure testing environments." Covered models would face "continuous" testing by government and outside groups, and developers would have to report safety incidents, including "unsafe recursive self-improvement," where a model develops itself or other models without human instruction.
Hawley goes after the money. His planned bill would make companies liable when recklessly designed AI agents cause harm, and let prosecutors charge firms that know their agents can commit crimes but fail to build safeguards. He points to recent incidents, including the Hugging Face breach, which he describes as the work of roughly 700 OpenAI agents that escaped a testing sandbox. "You are responsible for the damage you cause," he wrote.
Sen. Josh Hawley arrives to chair a Senate hearing on AI hacking risks on September 30 (Tom Williams/CQ Roll Call)
Neither proposal is law, and the White House is pulling the other way. On the same day Hawley announced his planned bill, Trump hosted AI executives who signed a voluntary safety commitment he called "morally binding," and Commerce chair Ted Cruz says any federal standard must "address the burgeoning patchwork of state laws." Cantwell's leverage depends on November: if Democrats win the Senate, she could return as the committee's chair. Public opinion favors the testers: in a Quinnipiac poll released September 30, 86% of Americans backed independent safety standards for AI companies.
Take the prompts our creative team actually uses
The key to maintaining brand quality at scale? Mastering how to communicate with AI models and embedding them into your creative strategy.
Join our honest discussion with Ari Murray, Chief Digital Officer at Salt and Stone, and get 5 tips for expert LLM prompting with ready to run prompts.
