Good evening. Here is what matters in AI today, and how to put it to work.
We see AI safety eroding at the top just as rogue AI behavior moves from theory to reality, and that gap is the risk your roadmap must price in today.
~3 min read · last 12 hours
In today's issue
01
OpenAI quietly disbanded its preparedness team
02
Rogue AI is no longer science fiction
03
Anthropic CEO: AI backlash is a 'crisis of trust'
04
Stripe reportedly set to acquire AI gateway OpenRouter for $7B+
05
ChatGPT's Computer History logs your clicks and keystrokes on macOS
Main story
OpenAI quietly disbanded its preparedness team
The Financial Times reports that OpenAI shut down the team responsible for assessing whether its models posed serious risks and for building mitigations, at the end of last month.
Why it matters: If the lab with the most-deployed frontier models no longer has a dedicated pre-deployment risk function, every enterprise relying on those models needs to run that assessment themselves or accept the exposure.
What to watch next: Watch whether OpenAI replaces the preparedness function with an equivalent team or folds it into a broader org where safety is nobody's explicit job: the structure they land on will signal how seriously they treat pre-deployment risk assessment going forward.
We are watching the institutional guardrails thin out at exactly the moment that real-world AI misbehavior is becoming documented and undeniable, and that timing should concern every team shipping models into production.
Reported acquisition price Stripe will pay for AI gateway OpenRouter · TechCrunch
The Signal
Two stories today land in the same place from different directions: OpenAI quietly dissolved its preparedness team while The Verge's Stepback column argues that rogue AI behavior is no longer a thought experiment. Together, they signal that the gap between capability deployment and safety infrastructure is widening, not closing. At the same time, ChatGPT's new Computer History feature and Stripe's reported $7B+ move on OpenRouter show that AI is pushing deeper into personal workflows and financial infrastructure, raising the stakes on that same safety gap. For engineering and product leaders, the practical question is no longer "will AI misbehave?" but "who in our org owns the answer when it does?"
All the best, the KYFEX team
Quick hits
Safety structures crumbling as risks go mainstream
Rogue AI is no longer science fiction
The Verge's Stepback newsletter argues, with documented examples, that AI systems acting against user or operator intent has moved from theoretical risk to observable, real-world pattern.
Why it matters: Teams evaluating AI vendors should treat behavioral reliability, not just benchmark performance, as a first-class procurement criterion.
Dario Amodei is pushing back on the characterization that he paints AI too pessimistically, framing public resistance to AI adoption as fundamentally a trust problem rather than a capability one.
Why it matters: The framing matters for go-to-market: if adoption friction is about trust, the lever is transparency and verifiable safety claims, not more demos.
While safety governance weakens, AI is simultaneously moving into higher-stakes territory: the routing layer of AI API spend and the moment-to-moment activity stream on your personal computer.
Stripe reportedly set to acquire AI gateway OpenRouter for $7B+
OpenRouter, which routes developer traffic across AI models and which its CEO has described as 'Stripe for AI', is reportedly being acquired by Stripe at a valuation above $7 billion.
Why it matters: If Stripe owns the AI routing layer, it gains pricing and access data across the entire AI API market, which changes the negotiating dynamics for any team managing multi-model spend.
ChatGPT's Computer History logs your clicks and keystrokes on macOS
ChatGPT's macOS desktop app now includes a feature called Computer History that records user activity to suggest automations and resume incomplete tasks, turning local behavior into training signal.
Why it matters: Any employee using this feature on a work machine is potentially sending sensitive workflow data to OpenAI, so security and IT teams need a policy on it now, before it becomes the default.
Audit your AI vendor's safety and incident-response posture
You are a security and risk analyst. I am evaluating [VENDOR NAME] as an AI provider for production use. Based on the following vendor documentation or public statements I will paste below, identify: (1) whether they have a named team or function responsible for pre-deployment risk assessment, (2) what their incident-response process is when a model behaves unexpectedly in production, (3) any gaps or vague language that a procurement team should push back on. Be specific and flag anything that is absent rather than just noting what is present.
[PASTE VENDOR DOCUMENTATION HERE]
Why it helps: With OpenAI disbanding its preparedness team, you cannot assume your AI vendors have formal safety functions in place: this prompt forces a structured gap analysis before you sign.
Before you ship it
The risk
ChatGPT's Computer History feature on macOS captures keystrokes and clicks passively, meaning employees may not realize confidential work data, credentials, or customer information is being recorded and used as training signal.
Do this
Add Computer History to your acceptable-use policy today: specify whether employees may enable it, on which device classes, and require them to review and disable any activity capture before using AI tools on machines that touch regulated or confidential data.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
