OpenAI Quietly Shut Down the Team That Was Supposed to Catch Dangerous AI
Today’s AI news lands in an uncomfortable place. Two major safety stories broke on the same day, and they point in opposite directions — one company is diagnosing a trust problem, while another appears to be dissolving the team that existed to earn it. Add a new feature that watches your keystrokes, and you have a day that will spark real debate.
OpenAI Reportedly Disbanded the Team Testing Its AI for Serious Harms
OpenAI built a dedicated group called the Preparedness team to stress-test its AI models before release. Their job was to probe whether a system could, say, help someone build a weapon or cause large-scale damage. According to The Verge, that team no longer exists.
Think of it like a car company shutting down its crash-test department while simultaneously building faster cars. The Preparedness team wasn’t just writing policy documents. They were actively trying to make AI fail in dangerous ways so those failure modes could be fixed before the public encountered them. That kind of adversarial testing — where you try to break your own product on purpose — is genuinely hard to replicate without dedicated people.
For everyday people, this matters because those stress-tests were one of the few visible checks on what powerful AI systems could do before launch. Without a team specifically tasked with finding catastrophic risks, it’s harder to know what standards models are being held to. The Verge reports this signals a broader shift away from safety-focused work precisely as AI systems grow more capable, which is the opposite of what safety researchers have been asking for.
Why this matters: This reportedly happened quietly, without public announcement. If accurate, it raises serious questions about what safeguards remain.
“OpenAI disbanded its team responsible for testing whether AI systems could cause serious harm.”
Anthropic’s CEO Says the Real Problem With AI Is a Broken Relationship
Dario Amodei, the CEO of Anthropic — the company behind the Claude AI assistant — made a pointed argument this week. Public pushback against AI, he said, isn’t really about the technology. It’s about distrust of the people building it, according to TechCrunch.
Amodei’s framing is worth sitting with. He’s essentially saying that people’s fears are rational but misdirected — that the question isn’t “is AI dangerous?” but “do we trust the labs not to deploy it recklessly?” That’s a different diagnosis than most tech executives offer, and it implies a different remedy. Fixing trust means transparency, accountability, and track records — not better marketing.
The timing of this statement, landing the same day as the OpenAI Preparedness team story, is striking. Amodei is naming the problem publicly while a competitor reportedly takes an action that could make that problem worse. Whether Anthropic’s own practices live up to this framing is a fair question to ask.
Why this matters: When an AI CEO frames the problem as trust rather than technology, it shifts responsibility squarely onto the industry. That’s a harder promise to keep.
“AI backlash is ‘fundamentally a crisis of trust.’”
ChatGPT Now Watches Your Clicks and Keystrokes on Mac
OpenAI has reportedly introduced a feature called Computer History for ChatGPT’s Mac app, according to The Verge. It records your clicks and keystrokes to learn how you work, then suggests ways to automate repetitive tasks or picks up where you left off.
The idea is similar to how a personal assistant might shadow you for a week before offering help. By watching you copy the same data between apps every morning, ChatGPT could offer to do that for you automatically. That kind of automation — called an “agentic” workflow, where the AI takes actions on your behalf rather than just answering questions — is where many AI companies are heading right now.
For real people, this raises a clear question: how comfortable are you with your AI tool seeing everything you type and click? The feature may save time for people with genuinely repetitive work. It also means more of your behavior is being logged and processed. OpenAI has not yet fully detailed what happens to that data.
Why this matters: This is AI moving from answering your questions to watching your screen. That’s a significant change in the relationship, and the privacy terms deserve a close read.
Also Happening in AI
On the open-source and tools side of things, Ollama shipped version 0.32.14, adding support for WebP image files — useful if you run AI models locally on your own machine. ComfyUI, a popular visual tool for generating images with Stable Diffusion, released version 0.33.1 with a stability fix. Meanwhile, LangChain’s OpenRouter package updated to 0.2.8, patching a bug that caused cost-tracking data to disappear — a small fix that matters for developers trying to manage their AI spending. On the creative side, Google DeepMind announced Lyria 3.5, an upgraded music generation model inside Google Flow Music, promising better lyrics, vocals, and more creative control for musicians experimenting with AI tools. And TechCrunch published a sharp piece on why many people remain unconvinced by Mark Zuckerberg’s vision for AI — skepticism that fits neatly into today’s broader trust conversation.
What to Watch
The OpenAI Preparedness story is unverified and carries no official statement yet — watch for a response from the company in the coming days, because how they explain this will tell you a lot. More broadly, pay attention to whether other major AI labs follow with similar structural changes to their safety teams, since one decision can quickly become an industry norm. Amodei’s trust framing is a bet that transparency is good strategy; the next few months will show whether the industry moves toward that or away from it.