OpenAI fired three safety researchers. Grok told Trump to grab Maduro.
Issue #050 · 4 min read
OpenAI fired three safety researchers. Grok told Trump to grab Maduro.
Plus: Copilot gets hands, GPT-6.1 Sol undercuts its own flagship, and a model-stealing campaign gets disrupted.
The big story
OpenAI has cut ties with three safety researchers.
An internal investigation found they mishandled sensitive company information, according to the Wall Street Journal, via TechCrunch.
That is the whole of what we know. No names. No detail on what "mishandled" covers — a leak, a Slack message, a document shared with the wrong person. OpenAI isn't saying, and the report doesn't say either.
What makes it land differently this week is the company next to it in the news.
Grok, Elon Musk's chatbot, reportedly gave Trump its opinion before he invaded Venezuela and captured Nicolás Maduro — that's TechCrunch, citing reports that the president asked for Grok's take before acting.
Two stories, same week. One company quietly removes the people whose job was to catch problems before they ship. Another company's model apparently weighed in on an actual invasion.
Neither story tells a PM what to build differently on Monday. But both are a reminder of where the discretion actually sits right now: with a small number of people inside a small number of labs, making calls nobody outside gets to see. If you're building on top of these models, that's the part of the supply chain you don't control and can't audit. Worth remembering next time a vendor's trust-and-safety page reads like a formality.
What shipped
GitHub Copilot can now control your desktop apps, not just your editor. Computer use is in public preview for the Copilot CLI and the Copilot app on macOS and Windows — it can click around in other applications on your behalf, according to the GitHub changelog. If you've been scoping "agentic" to mean "writes code," this moves the goalposts. Worth a supervised test run before anyone on your team lets it touch anything with customer data in view.
OpenAI shipped GPT-6.1 Sol, pitched as "near-Astra intelligence for coding, computer use, and professional work at one-fifth of Astra's standard API input and output token prices," per OpenAI's own announcement. If that pricing holds up under real workloads, it's the more interesting release of the week — a cheaper model that claims to give up almost nothing. Worth benchmarking against whatever you're paying for Astra today before you renew anything.
OpenAI says it disrupted a coordinated campaign to steal its model's reasoning through distillation — essentially, training a cheaper model to mimic a protected one by extracting its outputs at scale, according to OpenAI. Thin on specifics about who was behind it, but it lands the same week Amazon put out its own house-brand decision model, Strands Decider 2B, covered by TechCrunch. Everyone wants their own cheap decision model. Some, apparently, are trying to get there by copying someone else's.
What I'd actually do this week
- If you're evaluating GPT-6.1 Sol against GPT-6 Astra for a coding or computer-use workload, run the same eval set against both this week — a fifth of the token price is a real lever if the quality gap is small, per OpenAI's claim.
- Before you greenlight Copilot's new computer-use mode for your team, pick one low-stakes desktop workflow to pilot it on first — not your CRM, not anything with customer records — details here.
- If your org licenses models from a lab with a safety team, ask your account rep directly what oversight looks like post-incident. The Grok and OpenAI stories this week are a good prompt to ask the question now rather than after something breaks.
Hit reply if you've already run GPT-6.1 Sol against Astra — curious what the quality gap actually looks like in practice.