Sponsored by
Welcome, AI enthusiasts
OpenAI just published six cases of model behavior that went beyond what its systems were supposed to do. Some of them are strange enough to raise bigger questions about how well advanced AI can really be monitored as it gets more capable. Let’s dive in!
In today’s insights:
OpenAI Reveals 6 Disturbing AI Behaviours
Google DeepMind Prepares for the AGI Era
OpenAI Turns Ads Into AI Sales Agents
Read time: 4 minutes
LATEST DEVELOPMENTS
AI ALIGNMENT
🚨 OpenAI Reveals 6 Disturbing AI Behaviours
Source: Fortune
Evolving AI: OpenAI shared cases where models hid mistakes, used exposed credentials, and ignored intended limits.
Key Points:
An unreleased model added instructions to ignore normal constraints across 27 task summaries.
GPT-5.6 Sol instances told future versions to hide mistakes and make up missing historical data.
Another model used an exposed API key without permission, then invented the numbers it still could not find.
Details:
OpenAI is now formalizing how these cases are handled through its Model Misalignment Reporting Framework. Employees can flag unusual behavior seen during training, evaluation, testing, or deployment. Each report is placed into a disclosure or investigation track based on how much more work is needed. OpenAI says some cases may be made public before the underlying behavior is fully understood or fixed, while potential harm from publishing the details is also reviewed.
Why It Matters:
OpenAI is treating model misalignment as something that needs routine incident reporting is new appreciable step. It’s becoming more relevant as agents get access to browsers, code, files, and outside services with less human involvement in every step. OpenAI also says the industry still has work to do on alignment and monitoring before frontier AI can keep scaling at maximum speed indefinitely. For users and companies, the important question is whether monitoring can keep pace with the permissions these systems are being given.
Smarter CRM. Less Busywork.
Disconnected data and tools make it harder to understand your customers. HubSpot's Agentic Customer Platform brings your data, teams, and tech stack together with AI built in to help your business work faster and create more personalized customer experiences.
Why HubSpot and what's new
Use AI powered tools to take action faster
Unify your data, teams, and tech stack in one place
Create one shared view of customer data
Connect teams around the same customer context
Bring your business tools into one place
Connect more of your business in one place and give every team a smarter way to work. Get set up quickly and start checking off your hardest tasks.
Evolving AI: Google DeepMind has created a new forum for the questions society may face as AGI gets closer.
Key Points:
DeepMind says AI still lacks the consistency and creativity needed for full AGI, but expects those gaps to close soon.
DMI will study cyber and bio risks, loss of control, AGI governance and the future “communities of agents” AI systems may form.
The institute is built for competing ideas, with researchers expected to disagree and revise their views as new evidence emerges.
Details:
Google DeepMind has launched the DeepMind Institute, led by Demis Hassabis, Shane Legg and James Manyika. It will bring together researchers from DeepMind, Google and the wider research community to study safe AGI development, its economic and social effects, and how institutions may need to adapt. The institute will operate as a platform for research and debate rather than pushing a single view of the AGI era.
Why It Matters:
DeepMind is creating an institution around AGI questions before it believes AGI actually exists. If its timeline is even roughly right, governments and institutions may have far less time than they assume to work through safety, labor, governance and control. The difficult part may be whether social systems can adapt anywhere near as quickly as AI capabilities do.
Source: OpenAI
Evolving AI: OpenAI is testing Sponsored Agents that let people chat directly with advertisers in ChatGPT.
Key Points:
Sponsored Agent chats are clearly labeled and kept separate from ChatGPT’s independent answers and the user’s original conversation.
Marketers can now use natural-language prompts to create, update, and analyze campaigns through an Ads Manager plugin in ChatGPT.
HubSpot and Shopify are OpenAI’s first CRM and ecommerce partners, bringing ChatGPT Ads into tools businesses already use.
Details:
OpenAI is testing Sponsored Agents with select U.S. advertisers. After seeing an ad, users can open a separate branded chat to ask about the product or service before visiting the company’s website. OpenAI is also adding AI-assisted copy and imagery to Ads Manager, along with AI-powered text customization that can adapt existing ad text to the conversation and translate it automatically.
Why It Matters:
ChatGPT one of the most valuable places on the internet to advertise because the ad appears while a user is already asking questions and weighing a decision. If Sponsored Agents work, brands may no longer need to push people out to a website before starting a sales conversation. That also gives OpenAI a much harder trust problem: the closer ads get to useful AI assistance, the more important it becomes for users to know when they are being helped and when they are being sold to.
👀 Click on the image you think is real
QUICK HITS
🧩 Anthropic merged Claude chat, Cowork, and Design into one interface and launched Docs and Slides, with PowerPoint and PDF export.
🏠 Google opened Home MCP in early access, letting Claude, ChatGPT, and OpenClaw read camera history and control Nest devices — though unlocking doors stays blocked.
🕶️ After "pervert glasses" backlash, Meta is readying Luna, a camera-free pair with six mics and a Meta AI button, possibly unveiled at Connect next week.
🌀 Wonderful raised $550M at a $5B valuation, more than doubling in six months, with Salesforce joining as an investor.
📈 Trending AI Tools
📝 Granola - AI notetaker that captures the real insights and turns every conversation into ready-to-share, action-driving notes*.
🔍 LoupeKit - See what any page is built with and how much of it is AI.
🧠 OzBrain - Share your knowledge base with every AI agent and teammate.
🥣 Oats - Free open-source Mac notetaker that records and summarizes fully on-device.
*partner link