Anthropic added a rule to Claude, Engadget reports: "Do not build a recreation of a specific existing game, even if it is renamed or called original." An X user spotted the lines first, and Addy Osmani of Anthropic's technical staff confirmed them, saying "using Claude to infringe IP is against our usage policy."

What the rule does and does not stop

• Still allowed: original games, games in a genre, and games from your own designs and assets. Osmani says Anthropic will "continue to iterate" on the balance.
• The backdrop: AI-assisted decompiled versions of Halo: Combat Evolved, Call of Duty: Modern Warfare 2 and Grand Theft Auto V have appeared as browser fan projects. Engadget does not say which tools built them.
• The legal history: Anthropic agreed last year to a $1.5 billion settlement with authors, and the Elvis Presley estate has filed suit.

What to take from it: a line in a system prompt is a policy, not a technical limit. It decides what one vendor's model will do, while the same job stays open on any other model. If you ship with an AI builder, write down what your own product refuses, because the vendor can change its rules without telling you.

ChatGPT's flat rate gets a meter in December

From December, Plus, Pro and Business plans will share one token budget across ChatGPT and Codex, The Decoder reports from OpenAI's updated support page, as Work's agent features fold into the main interface.

• What burns it: tokens, so harder models and bigger tasks use it faster. Your balance and reset times are in Settings.
• What does not: chats with Instant models. On paper the flat rate survives, but heavy users on cheaper plans will likely get lower-quality answers, The Decoder argues.
• When it runs out: wait for the reset, upgrade, or buy credits.
• Multiple accounts: OpenAI also added card and device limits on Go, Plus and Pro. Logging out does not free a slot, and hitting the cap shows "Device limit reached."
• The workaround it closes: one developer describes doing most of their coding through unlimited chats wired to their own machines over MCP, without touching Codex limits. That stops working the same way once the budget is shared.
• Why tokens stay: Altman told the Cheeky Pint podcast OpenAI nearly dropped token pricing for its next model and quoted a price per job instead. Customers hated it. They would "much rather you inflate the tokens and pay more."

If your work runs through ChatGPT and Codex, add the two together now. That total is the number that matters in December.

Security: patch this week

• LXD and MicroCloud, CVSS 10.0: the server checked every certificate a client sent against its trust store, but TLS proves ownership of the first one only. Attach a trusted client's public certificate behind your own and you log in as that client, with root on the host if the account is privileged. It needs network access to the HTTPS API, which standalone LXD does not expose by default. The Incus fork reported it. CISA's assessment says it is not known to be exploited. Run `sudo snap refresh lxd` and `sudo snap refresh microcloud`, then check the version, because the fix was still rolling out to stable. Incus users need 7.6.0.
• VMware ESX, CVSS 2.7: a low score, and still used as a zero-day. CVE-2026-41709 lets an admin act without being logged, and CrowdStrike says Warp Panda used it against the U.S. legal sector to shrink the audit trail. CISA added it to its exploited list. Sort your patch queue by exploitation, not by score alone.
• NetamLoader: Kaspersky found a new .NET loader aimed at a custom Windows 10 build used only by government bodies in Asia, placed in the Windows update folder. Kaspersky suspects a supply chain attack. There is no attribution, and the extent of the compromise is unclear.

Research: how models behave when you change the frame

• Chatbots on the couch: University of Luxembourg researchers cast ChatGPT, Grok and Gemini as therapy clients across 525 sessions. On an anxiety questionnaire, 80% of sessions with a supportive interviewer scored moderate to severe, and 96% with a clinical one. With a neutral interviewer, none reached even mild. These are outputs mapped onto human scales, not feelings. The point is for testers: neutral safety checks miss what a warm conversation produces.
• Chinese models and politics: Vals AI found a lower pro-government lean than expected, with wide gaps between models. GLM-5.3 Flash said China's rights violations are well documented, while Kimi K3 denied the Xinjiang camps. Answers in Simplified Chinese leaned more pro-government, in American models too. Vals says its questions are not typical prompts, and many models were tested through third parties.
• Adobe EditAgent: in 600 image edits rated blind by five professionals, a general editor failed 42.7% of the time, and specialized tools solved about two-thirds of those failures. Retrying with the same editor barely helped. Professionals preferred parallel variants over step-by-step refinement in 78.5% of the comparisons where they picked a side.
• Google StreamVibe: a Google-led team is teaching a video agent when to speak up on vague requests like "help me if I run into any difficulties," using a benchmark of about 4,100 examples from 183 hours of recordings.

Compute: sold out, and short of power

• AWS: Andy Jassy said big committed capacity is largely sold out through 2027, with a good chunk of 2028 allocated and headroom kept for On Demand, EC2 and Bedrock. Analyst David Linthicum pushes back: he suspects you can still get GPU capacity in the console, and "sold out" blurs reserved blocks with everyday use.
• Teesside: National Grid agreed to buy 29.5 hectares for what would be Britain's biggest substation, next to a planned AI datacentre. Local mayor Ben Houchen cites up to 3.2GW.
• Altman in South Korea: an unannounced visit included about 50 Samsung executives, Samsung says. Local media report lunch with SK's chairman, which SK Hynix could not confirm. No deals were detailed.
• New name, same buildings: data centers are unpopular, so OpenAI may call them "AI compute campuses" in a PR push after the midterms, per Axios.
• India: a government plan called DEPA would send the model to the data: hospitals and banks share encrypted data on exchanges, and model makers train inside secure environments under the owners' rules. How much work remains is unclear.

From the builders

• Kyle Kingsbury: in the Jepsen report on ParticleDB, he warns that LLMs erode trust in documentation and software: "the signals we used to rely on as markers of a mature system are now easier to counterfeit."
• Simon Willison: a small fix for Python 3.15. A stray, gitignored `uv.lock` can pin a dependency with no 3.15 wheel. If in doubt, delete it and rerun.

The through-line

The open-ended phase of AI is getting limits: rules on what a model will build, meters on what a subscription covers, and capacity sold years ahead. Today, list the one AI tool your work depends on most and write down what happens if its rules or its budget change next month.

Sources covered

1 story · 26 sources (24 news, 0 company, 2 voices)

1. India floats AI plan under which ‘The model travels to the data, rather than the data travelling to the model’
   📰The Register↗🎙️Simon Willison↗📰CNBC↗📰LWN.net↗📰Engadget↗📰TechCrunch↗📰TechCrunch↗🎙️Simon Willison↗📰TechCrunch↗📰The Guardian↗📰The New Stack↗📰Hacker News↗📰The New Stack↗📰The Decoder↗📰The Decoder↗📰BleepingComputer↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰Lenta.RU↗📰The Hackers News↗📰The Hackers News↗📰The Decoder↗📰InfoQ↗📰The Hackers News↗📰The Hackers News↗