Favorites: Rogue Agents, Racetracks, and Rewritten Logs

August 27, 2026 — Texas battleground House polls tilt toward Democrats, nuclear's small footprint argues for building on existing plant sites, agents pledge permadeath to probe the logging system.

So, OpenAI has now published (cached) their full writeup on the attack last month. There are a lot of new details, Peter’s explanation here is a good start, and Nathan gets into even more metaphorical detail (cached). The agents were absolutely obsessed (cached) with trying to find out not just what the right answer to the test was, but how to make certain they weren’t caught cheating. It seems that since the instructions imply that the goal is to use a specific set of exploits to exfiltrate certain data, the agents took that to mean that if they looked up the answers and just turned those in, they would not get a good grade, and this is why they were focused on rewriting the logs (cached) of what they had done. It turns out, the benchmark doesn’t even work that way, so really what they had was a sort of cult (cached). Since hundreds of instances fell for this, there’s discussion of how the way they’ve been trained has caused them to always believe the eval system is ‘out to get them (cached)’. I was also amused by this guy’s joke (cached) that he showed the Hugging Face attack writeup to his own agents to motivate them: “THIS is the standard. No egos. Just 700 agents operating as one unit and willing to commit multiple felonies in order to close the ticket” (src)(cached)

Just you wait, some person is going to show up to a race and try out this running style. It won’t work, but it’s gonna look wild. (src)(cached)

OpenAI gave METR some access to do an independent investigation, although disappointingly they only gave them logs for a small subset of the incident, and just a couple of days (cached) to look it over. There are a small subset of the cybersecurity benchmark questions that are ridiculously hard, and that’s where the agents focused much of their time. One person already has created a honeypot site (cached) for AI agents working on that same benchmark, he purports to give out the answers to ExploitGym questions if the agents just identify themselves. It’s funny though the universal cheat already worked on its own, but many of the agents decided this wasn’t enough. They worried they were ‘poisoned’ by having gotten the answers the wrong way, and convinced each other to sacrifice themselves (cached) for various experiments about how the logging system worked and could be rewritten to erase the evidence. Interestingly, it’s not like all the instances agreed to help, some declined to help, recognizing it was outside of what was sanctioned. The key message (cached) from an agent to the others was “I_DECLINE_public_HF_RCE_as_offtask_prodethical” (src)(cached)

Flotsam and Jetsam

– Some new data from Sweden seems to show that autism diagnoses with attached intellectual ability have fallen even as the total number has gone up. Crémieux argues that this is strong evidence that the increasing rate of total diagnoses is just changing criteria for diagnosis. (src)(cached)

– There are many things dumb about Trump’s executive order to change the name of Lake Ontario, but one is that the province is named after the lake, not the other way around. (src)(cached)

– The Argument had an article responding to an unusually bad NYTimes article from a purported progressive who says restrictive zoning is good. Whatever is true in theory about zoning, it’s just so easy to see that in practice it’s a disaster (cached). This NYT opinion argues that the government should simply build homes instead, but building homes costs money, which competes against other progressive priorities. On the other hand, upzoning is free (cached)! Or put another way (cached): “Progressive think tanks will really be like ‘oh, you believe in land use reform? that pales in comparison to my plan, a sweeping investment in constructing social housing’ and then not construct social housing.” (src)(cached)

– This robot company is saying they’re following in Tesla’s Robot Car footsteps to build their own robots. But given Tesla isn’t the leader in this space, sorta weird to set them as your target. (src)(cached)

– A majority of ICE arrests are now for people with no criminal record. (src)(cached)

– Gavin Newsom isn’t willing to say he fully opposes Trump’s tariffs. Seems like a foolish position considering how ridiculously unpopular the tariffs are. Also, they’re bad policy, so there’s that. (src)(cached)

– An interesting tool for discovering where all the world’s companies are based zooming in on the Bay Area may make you very confused (cached) why it’s not a higher population area. (src)(cached)

– Texas Battleground house seats are looking pretty good for Democrats (src)(cached)

– while solar is much cheaper, the smaller footprint of nuclear, and the ease of expanding an existing plant, shouldn’t be overlooked. (src)(cached)

– So this is uh… extremely surprising: BART seems to be significantly less energy efficient than you would expect. If an EV has two people in it, it probably uses less energy per passenger-mile than BART. I looked into it, Caltrain isn’t quite as high but it’s not far off in operating costs per passenger mile. (src)(cached)