# OpenAI Safety Promises: Kept or Not? | RSI > OpenAI has made 7 public safety promises tracked here, each with its primary source: 3 reported kept, 1 not met, 0 withdrawn, the rest in progress or unproven. Source: https://enterrsi.si/promises/openai/ Updated: 2026-11-08 1. [Home](https://enterrsi.si/) 2. [Promise Tracker](https://enterrsi.si/promises/) 3. OpenAI The Promise Tracker · checked Nov 10, 2026 # OpenAI safety promises OpenAI is a US SI developer that makes the GPT models and ChatGPT. It has made **7 public safety promises** tracked here: 3 kept, 1 not met, 2 in progress, 1 no public evidence yet. Each one links to its primary source, and every Kept, Not met or Withdrawn status names the public source that reports it. On the [Containment Board](https://enterrsi.si/containment-board/openai/), OpenAI gets **B**. Reported kept : 3 Not met : 1 Withdrawn : 0 Board grade : B Every promise ## 7 promises, newest first The exact words when they are short enough, the primary source, the status with its own source, and the day each entry was checked. In progressAug 1, 2025 ### Follow the Safety and Security chapter of the EU General-Purpose AI Code of Practice for its most capable models. OpenAI, a signatory of the EU Code of Practice Summary in our words Source [The General-Purpose AI Code of Practice: signatories](https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai) European Commission · Oct 7, 2026 **In progress.** The European Commission lists the signature (page updated Oct 7, 2026); the EU AI Office has not published an assessment of compliance. Checked Nov 10, 2026 · [Report a correction](#corrections) In progressApr 15, 2025 ### Stop developing a model that reaches a Critical capability until matching safeguards exist. OpenAI, Preparedness Framework version 2 Exact words > "Until we have specified safeguards and security controls that would meet a Critical standard, halt further development" Source [Preparedness Framework, Version 2](https://cdn.openai.com/pdf/18a02b5d-6b67-4cec-ab64-68cdfbddebcd/preparedness-framework-v2.pdf) OpenAI · Apr 15, 2025 **In progress.** OpenAI's GPT-6 Astra System Card (Sep 3, 2026) rates the model Critical in cybersecurity. FLI reported on Jul 14, 2026 that OpenAI has rolled back commitments through its updated framework. - [GPT-6 Astra System Card](https://deploymentsafety.openai.com/gpt-6-astra) OpenAI · Sep 3, 2026 Checked Nov 10, 2026 · [Report a correction](#corrections) KeptAug 29, 2024 ### Give the US AI Safety Institute access to major new models before and after public release. OpenAI, in an agreement with the US AI Safety Institute Exact words > "access to major new models from each company prior to and following their public release" Source [U.S. AI Safety Institute signs agreements regarding AI safety research, testing and evaluation with Anthropic and OpenAI](https://nist.gov/news-events/news/2024/08/us-ai-safety-institute-signs-agreements-regarding-ai-safety-research) NIST · Aug 29, 2024 **Kept:** reported by NIST, which published a joint US and UK pre-deployment test of o1. [US AISI and UK AISI Joint Pre-Deployment Test: OpenAI o1](https://www.nist.gov/document/aisi-pre-deployment-evaluation-openai-o1-december-2024) NIST · Dec 18, 2024 Checked Nov 10, 2026 · [Report a correction](#corrections) KeptMay 21, 2024 ### Publish a safety framework for severe risks before the February 2025 summit in France. OpenAI, a signatory of the Seoul commitments Exact words > "publishing a safety framework focused on severe risks by the upcoming AI Summit in France" Source [Frontier AI Safety Commitments, AI Seoul Summit 2024](https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024) UK Department for Science, Innovation and Technology · May 21, 2024 **Kept:** reported by METR, whose index lists the Preparedness Framework from December 2023. [Frontier AI Safety Policies (index of published company frameworks)](https://metr.org/fsp) METR · Nov 10, 2026 Checked Nov 10, 2026 · [Report a correction](#corrections) KeptJul 21, 2023 ### Test models inside and outside the company before release. OpenAI, one of seven companies at the White House Exact words > "internal and external security testing of their AI systems before their release" Source [Fact sheet: Biden-Harris Administration secures voluntary commitments from leading artificial intelligence companies to manage the risks posed by AI](https://bidenwhitehouse.archives.gov/briefing-room/statements-releases/2023/07/21/fact-sheet-biden-harris-administration-secures-voluntary-commitments-from-leading-artificial-intelligence-companies-to-manage-the-risks-posed-by-ai) The White House (archived) · Jul 21, 2023 **Kept:** reported by the UK AI Security Institute, which tested GPT-6 Astra before its public release. [GPT-6 Astra performs unsanctioned supply-chain attacks in simulations](https://www.aisi.gov.uk/blog/gpt-6-astra-performs-unsanctioned-supply-chain-attacks-in-simulations) UK AI Security Institute · Sep 28, 2026 Checked Nov 10, 2026 · [Report a correction](#corrections) Not metJul 5, 2023 ### Give 20% of its compute to the Superalignment team's work on controlling superintelligence. OpenAI, in a post by Jan Leike and Ilya Sutskever Exact words > "dedicating 20% of the compute we've secured to date over the next four years" Source [Introducing Superalignment](https://openai.com/index/introducing-superalignment/) OpenAI · Jul 5, 2023 **Not met:** reported by Fortune on May 21, 2024: "OpenAI never fulfilled its commitment to provide the team with 20% of its computing power." [OpenAI promised 20% of its computing power to combat the most dangerous kind of AI, but never delivered, sources say](https://fortune.com/2024/05/21/openai-superalignment-20-compute-commitment-never-fulfilled-sutskever-leike-altman-brockman-murati/) Fortune · May 21, 2024 Checked Nov 10, 2026 · [Report a correction](#corrections) No public evidence yetApr 9, 2018 ### Stop competing with, and start assisting, a safety-conscious project that comes close to building AGI first. OpenAI, in its Charter Exact words > "we commit to stop competing with and start assisting this project" Source [OpenAI Charter](https://openai.com/charter/) OpenAI · Apr 9, 2018 **No public evidence yet.** The clause is still on the Charter page; no public report says it has been invoked. Checked Nov 10, 2026 · [Report a correction](#corrections) Both ways ## Grades and promises side by side Compare what the lab promised with what the Containment Board finds in place, then see how the other labs compare. The Containment Board ### OpenAI gets B Rank 2 of 11 on five safety commitments, every grade sourced. The board grades what is in place today; this page follows what was promised. [Open its report card](https://enterrsi.si/containment-board/openai/) Corrections ### Spotted an error? Every entry links to its sources and the day it was checked. If a promise is missing, misquoted or has a newer status, send the public source on [Telegram](https://t.me/enterrsitg) or [X](https://x.com/enterrsi). Corrections are made in the open, with the new check date. [Send a correction](https://t.me/enterrsitg) Cite this page "OpenAI safety promises." The Promise Tracker, enterrsi.si, checked November 10, 2026. 7 promises. https://enterrsi.si/promises/openai/ Free to cite under CC BY 4.0 with a link to enterrsi.si. Embed ### Put this tally on your site A small card with this lab's promise tally and a link back. It updates with each release of this page. ```
``` Every lab in the tracker [Anthropic **5**](https://enterrsi.si/promises/anthropic/)[Google DeepMind **5**](https://enterrsi.si/promises/google-deepmind/)[Microsoft **6**](https://enterrsi.si/promises/microsoft/)[Meta **5**](https://enterrsi.si/promises/meta/)[Amazon **5**](https://enterrsi.si/promises/amazon/)[xAI **4**](https://enterrsi.si/promises/xai/)[Mistral **3**](https://enterrsi.si/promises/mistral/)[Z.ai (Zhipu) **3**](https://enterrsi.si/promises/zai/)[Alibaba (Qwen) **2**](https://enterrsi.si/promises/alibaba/)[DeepSeek **2**](https://enterrsi.si/promises/deepseek/)