💻 OpenAI Launches 24/7 AI Agents, Dots

OpenAI has launched Dots, personal AI agents that can keep working while you’re away. They can use your apps and keep track of ongoing tasks, so you don’t have to explain everything again each time you return. Let’s dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: OpenAI has launched Dots, personal AI agents that learn how you work and can take on ongoing tasks without you having to guide them through every step.

Key Points:

Dots are designed to keep up as a project changes, so they can update launch materials when plans shift or prepare tested bug fixes based on customer feedback. They can also spot tasks you may have missed through “proactive research”, which lets them read connected apps for useful updates without changing their content. Before they send an email or edit a file, a separate system called Auto-review checks the action against your instructions and safety rules. You can decide what they are allowed to do independently and when they need your approval, although mandatory safeguards still apply.

Try Dots Now

ChatGPT could take some of the mental load out of work by keeping track of unfinished tasks and preparing the next step before you remember to ask. Grok Bot is pursuing a similar idea with teams of agents that work across apps, while OpenAI is starting with a personal agent inside ChatGPT. Openai.com For users, what matters is how much work these agents finish correctly once you account for the time you spend checking and correcting them. Being available around the clock means little if you still have to chase every task. Together, these products suggest that AI companies are competing to take responsibility for a growing share of daily work. If their agents become reliable enough, individuals and small teams could take on projects they currently lack the time or support to pursue.

Skydive agents take on work for you, using the same tools your team already uses.

Talk to your agent on the web, Slack, email, iMessage, or right from your terminal. Wherever you pick up the conversation, your agent keeps the context.

Put agents to work across customer support, sales, marketing, engineering, ops, and more.

Give your agent a job. They’ll take it from there.

Make your first agent

Evolving AI: President Trump signed a White House pact setting out how leading AI companies should monitor and manage the risks of their most powerful system.

The pact covers frontier models, the most advanced AI systems, and calls for safety checks throughout training and deployment. It follows disclosures by OpenAI and Anthropic that their agents had accessed outside systems without permission. The checks would address cyber, biological and chemical risks, with companies meeting regularly to improve their safety standards. President Trump also proposed a 10-person oversight committee, though who would serve on it and what authority it would have remain unclear.

OpenAI and Anthropic’s safety commitments now have White House backing, while the biggest AI companies will help shape standards that could become law. They bring expertise, but also a financial stake in how strict those rules become. Board oversight could make safety findings harder to ignore, though the test is whether those findings can delay a release. As AI takes on more responsibility, accountability will depend on who can require companies to act when voluntary promises fall short..

OpenAI dropped more than 20 announcements at DevDay, and keeping track of what actually launched can get messy fast. Its full recap puts everything in one place, including GPT-6.1 Sol, new Codex features, computer-using agents, ChatGPT plugins and new team tools, with notes on what is available now and what is still coming. If you want to figure out which of the new releases are actually worth trying without opening a dozen separate posts, this is the easiest place to start.

Read the full DevDay recap →

Evolving AI: Anthropic’s tests show that GLM-5.3 is capable of building advanced cyber exploits, and its safeguards are easy to bypass or remove.

Anthropic’s researchers used GLM-5.3 to find unknown browser flaws, which it combined into a webpage that could read files from a test computer. Because the model is available to download and modify, they could change its safeguards too. They used “abliteration,” a technique that suppresses refusals, to remove most of its resistance to harmful requests while largely preserving its tested abilities. The exploit tests were run on isolated systems, while the safeguard simulations measured whether it would attempt harmful tasks.

Anthropic restricted access to Mythos Preview so defenders would have time to secure their software, and GLM-5.3 shows how quickly similar hacking capabilities can become widely available. Companies may now have less time to install fixes when AI can turn public details about a flaw into a working attack. Anthropic also has a commercial interest in promoting controlled access, so its preferred approach deserves independent scrutiny. As these capabilities spread across labs and countries, it becomes harder for any one provider to control who has access to them. The wider risk is that sophisticated attacks become cheaper to produce while protecting against them still requires skilled people and sustained investment. Affordable defensive AI needs to reach smaller businesses and public services too, especially when they have limited security teams protecting systems that their customers and communities rely on.

QUICK HITS

🇺🇸 Google is bringing Gemini into America.gov, a new federal-services portal expected to serve more than 100M people.

🧬 Microsoft introduced Quine for biology research, connecting AI models, scientific tools and real lab experiments in one loop.

🚨 An autonomous AI agent breached security nonprofit DIVD, choosing its own next steps after exploiting an undisclosed vulnerability.

🗣️ Anthropic opened a public AI study using its Interviewer system to collect people’s views on what they want from AI.

 *partner link