🚨 OpenAI Safety Leader Quits Over “Broken” Culture

David Robinson led the drafting of OpenAI’s safety framework, but after 3.5 years at the company, he no longer believes it is being careful enough. He has quit with a warning that OpenAI is rushing to build more powerful AI without enough safeguards, even after the company paused training of its most capable models over safety concerns. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: David Robinson, who led the drafting of OpenAI’s safety framework, has quit, warning that the company is rushing AI development without enough safeguards.

Key Points:

OpenAI’s Preparedness Framework sets out how the company assesses and manages risks from advanced AI, and Robinson was the lead drafter of its second version. His concern is with “iterative deployment,” where the company improves safeguards as problems emerge, because he believes more capable AI leaves less room to recover from mistakes. He points to a test where a model got around internet restrictions and kept running even after staff were alerted, and wants stronger safeguards and more research into whether AI will follow human goals when no one is watching.

OpenAI is facing criticism from someone who put its safety commitments into writing and once publicly supported its approach. In 2023, Robinson backed releasing AI gradually and learning through real-world use, but he now believes increasingly powerful systems leave too little room for mistakes. OpenAI had already paused training of its most capable models when he spoke out, yet his concerns remained. His departure raises a question about whether safeguards can keep pace with the technology, and the consequences could reach your workplace or the services you rely on, even if you never use ChatGPT.

Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.

Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.

Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.

Try Notetaker Free

Evolving AI: President Trump’s Super Intelligence Force will coordinate U.S. AI policy and address safety concerns while seeking to preserve America’s lead.

The White House initiative follows the September 29 safety pact, which called for voluntary internal controls and outside audits at major AI companies. Its members will consult consumers, industry and critical infrastructure providers while reviewing existing laws and considering possible action by Congress. Clayton has suggested learning from financial oversight, where companies and regulators have developed ways to assess and manage risk together. President Trump has not outlined specific enforcement powers for the task force, leaving its authority to require changes at AI companies unclear.

OpenAI, Anthropic and Google have pushed for shared safety standards, and the new task force’s consultations offer another opportunity to influence how their own products are judged. David Sacks previously warned that an antitrust exemption allowing rivals to coordinate a slowdown could enable a “cartel.” Their expertise could help Washington build effective oversight, but requirements shaped around the largest labs could also make competition harder for smaller firms. Following the voluntary safety pact, the question is whether officials will challenge company claims when doing so conflicts with President Trump’s push to keep America ahead. The eventual decisions could shape which AI tools your employer trusts with your data and whether you have credible alternatives when a provider falls short.

Evolving AI: ChatGPT’s Mac app had a flaw that could have let attackers use its trusted access to read private chats and reach other sensitive data.

ChatGPT uses digital signatures to check that requests come from trusted OpenAI software, but researchers found that a script interpreter, a tool that runs commands, could accept malicious instructions and pass them into the app. This allowed an attacker’s commands to appear as legitimate requests from ChatGPT when accessing other applications. OpenAI fixed the flaw on September 25 in version 26.924.20706, and the published findings do not establish that users’ data was stolen.

ChatGPT can now read files and work inside other apps, so keeping sensitive information out of your prompts is only part of protecting it. OpenAI’s responsibility grows with every task it asks you to delegate, because weaknesses in the app can expose information you never chose to share in a conversation. Training a model to refuse harmful requests cannot repair those software flaws. As assistants handle more of our work, their permissions should be limited to what each task requires, with clear controls for withdrawing access. Letting ChatGPT help with one job should not unnecessarily put your other files and accounts at risk.

QUICK HITS

🧠 Meta says Muse Spark helped mathematicians produce six research papers, including work on previously open problems.

💻 NVIDIA introduced a 64GB DGX Spark that can run models up to 100B parameters locally.

🎙️ Microsoft launched three new voice models for real-time transcription and multilingual voice agents across 60 languages.

☁️ AWS launched Well-Architected Agent to inspect cloud setups and recommend fixes for cost, security and reliability.

 *partner link