OpenAI Slows Down and Focuses on Safety - DTNS 5335

**Tom Merritt** (0:08)
This is the Daily Tech News for Wednesday, August 19th, 2026 We tell you what you need to know, give you important context, and gosh darn it, try to help each other understand.

**Sarah Lane** (0:17)
Today, OpenAI is announcing new plans to keep you safe from those rogue agents.

**Tom Merritt** (0:23)
Coming from everywhere. I am not a rogue agent, I'm Tom Merritt.

**Sarah Lane** (0:27)
I'm coming from inside the house, and I'm Sarah Lane.

**Tom Merritt** (0:30)
You literally are.
Let's start with what you need to know with the big story.
Yeah, so OpenAI made a post, which is the beginning, I suppose, of disclosing changes to its safety practices. We had some rumors recently that they had changed their safety team to kind of move safety people into the actual production teams. And this seems to bear that out if you think about it that way. Recently, OpenAI announced that it would pause development on an internal model it calls Astra for two weeks because they believed Astra had reached a critical threshold for security capabilities. OpenAI now says it is pausing development on more models that are farther from shipping than Astra. I'll just quote from the posting. Our largest planned Frontier RL reinforcement learning run remains on hold while we conduct smaller scale training and evaluations to assess model behavior, validate our safeguards, and establish more evidence of alignment before proceeding. So what that means is OpenAI is always training new models for the most part. It is not training those new models right now while it figures out its new safety practices.
OpenAI is also rewriting its preparedness framework. The last time it updated it was back in 2023 It is strengthening monitoring in its development process by introducing alignment. Alignment is what the model is trying to achieve the same as what humans want it to try to achieve. So they're introducing alignment and security safeguards earlier in the development of models, and also adding some higher safeguards post-training. So adding the safeguards in earlier and then raising them after you're done training when you're doing the inferencing and testing. OpenAI will also add controls on workloads from the internet during testing. This was a big factor in the rogue agent stuff.
They wrote, we have designed these controls so that a single compromise of a workload or a supporting service does not by itself allow for unauthorized access to the internet or other internal networks. That's a hindsight is 2020 move, but a good one. It will remove potentially vulnerable shared services. So not just vulnerable shared services, but ones where it thinks like maybe they could find a bug there. We're going to just pull that one out there and also do some simulated attacks to test boundary conditions. Issues will be flagged within 30 minutes of detection. In other words, they shouldn't be more, the faster, the better, but no more than 30 minutes. And security and research teams should respond immediately once they're flagged. And if they can't identify it as a positive in 30 minutes, then pause the model activity. Don't wait to keep looking. It will monitor all models in training. It didn't always do that. It only monitored the most high risk ones. But now it's going to monitor all the models in training. And it will also monitor inferencing. That's when you ask at a query and it kicks back after training.
If a model displays critical capabilities, like Astra. OpenAI noted that some security measures will increase its compute overhead by 20% of the observed inference workload, but said that's for internal research, so it's not going to impact how much you pay for the API. It will raise the overall cost of running OpenAI, so it's going to impact cost down the road. But it's not something that's raising the compute for you on the other end. This is all for the internal research. The company is increasing the computer resources for understanding the way systems reason and act. They wrote, those systems examine tool actions, available reasoning, and the full sequence of activity for unauthorized access, data theft, destructive behavior, and attempts to defeat safeguards.
And that's just the beginning. That's what they wrote today. OpenAI says it's going to share more details about its monitoring scheme in a future post. Sarah, do you feel more secure?

**Sarah Lane** (4:30)
Yeah, I do. I feel like if you are somebody who is suspicious of OpenAI or any company that is building models like this, then you might be suspicious of what the company has said. Here are all the safeguards that we're putting into, the chain of command.
And I think that OpenAI being extremely transparent about this is the right call. Let folks know what happened, why you realize that other safeguards need to be put into place, and let everybody know exactly how that works. Because when nobody knows how it works, that's where a rogue agent jumps out of the sandbox, does something weird, and then we all talk about it for three weeks.