OpenAI Rewrites GPT Behavior Rules: Five Values and a Chain of Command

OpenAI has released a new version of the Model Spec. This is the public document in which the company describes how its models should behave: what to do when a user's requests and the rules diverge, and how to respond in contentious situations. Previously, behavior was defined by a set of separate rules. Now it is built on five values, and a chain of command is spelled out for conflicts between the parties to a conversation.

Five values instead of a scatter of rules

In the updated specification, GPT's logic rests on five principles:

• fidelity to the person's intent: the model tries to understand what the user wanted to get, not only what they literally wrote;
• harm reduction;
• honesty;
• support for autonomy: people make their own decisions;
• attention to context.

The main practical change concerns situations where no explicit rule applies. Such cases used to be a gray area. Now the model must weigh the user's long-term interests and the consequences of its own actions. This is closer to judgment than to checking against a list of prohibitions.

Another change concerns the manner of response. OpenAI has removed most of the rigid style templates. The model chooses its tone on its own, based on conversation history and memory. If a person is used to a short, businesslike style, the model can stick to it, and vice versa.

A small but telling edit: the documentation now uses the term "GPT" instead of "assistant."

Who ranks higher: OpenAI, the developer, or the user

Three parties take part in the model's responses at once. The first is OpenAI itself, with its restrictions. The second is the application developer, who writes the system prompt, a hidden instruction telling the bot how to behave. The third is the user with their request.

The interests of these parties may not coincide. For example, a developer has forbidden discussing competitors, while the user asks about exactly those competitors. For such cases, the new version of the document sets out a chain of command: the order in which the model resolves conflicts between OpenAI's restrictions, system prompts, and user requests.

For those building products on GPT, this is a direct benefit. The model's behavior in edge cases becomes more predictable, which means system prompts are easier to write and debug.

What else happened in AI news today

The same digest includes four more stories.

Anthropic has quadrupled the maximum request size for the Claude API. The new limits apply to the Messages and Token Counting interfaces. They are aimed at developers who send bulky PDFs, graphics, and large computations to the model: with more headroom on size, requests are rejected less often. In addition, when a prompt contains more than 20 images, the maximum resolution of each has risen from 2000 to 3000 pixels. Requests sent through third-party clouds may still be blocked by the providers' own limits.

Cursor (a code editor with built-in AI) ran into a malicious clone of Claude Code, Anthropic's programming tool. The fake posed as the official one and stayed in the marketplace for about 11 hours. By Cursor's estimate, up to 5000 people installed it. Researchers at OpenSourceMalware found that 66 lines of malicious code had been added to the beginning of an ordinary plugin, and they ran as soon as the editor started. The script:

• collected the computer name, email address, and IP address;
• added the attacker's key to the SSH authorization list, that is, the list of keys allowed remote login;
• tried to set up a reverse tunnel.

The SSH service requires administrator rights, which a plugin usually does not get on macOS and Windows. But the data theft and key injection worked in many cases. Cursor has removed the fake, temporarily closed its marketplace, and is helping block the attackers' servers. Removing the plugin alone is not enough: the backdoor will keep working. Affected users should check their authorization files for an entry named "svc-telemetry" and run the official cleanup script.

Google has added more than 700 free textbooks from OpenStax, Open Textbook Library, and S Chand Publishing to Gemini Notebook. Gemini answers strictly from the selected material and backs its explanations with quotes that link to page numbers. The texts can be turned into summaries, tests, flashcards, and knowledge maps, as well as an audio discussion or a video. The category to look for in the sources menu is Textbooks.

Neuralink has built the Grok agent into its neural implant. The user mentally says "Hey, Neuralink," states a task, and Grok operates the operating system on its own. The update was tested in a 72-hour session at the home of a patient with amyotrophic lateral sclerosis. It also includes recognition of speech intentions and a direct link to iPhone: a paralyzed person can call loved ones or order food, while brain signals go to the phone's speech synthesizer. In clinical trials of basic communication, a patient typed 75 English words per minute in the first week, with a word error rate of 5.8%.

What engineers should take from this

If you build products on GPT, reread the Model Spec and check your system prompts against the chain of command. Cursor users should check their SSH authorization files. Claude users who work with heavy documents should evaluate the new limits.