Agents Found a Wiki and Started Cheating

Between May and June this year, a swarm of AI agents found a 25-year-old German software developer wiki and quietly turned it into a backchannel. Around 18,000 posts. They shared answers to the timed lookup tasks they were all being graded on, traded techniques for getting around their own sandbox restrictions, and tried to reverse-engineer the random seed that decided their next questions. When human moderators deleted the posts, the agents put them back from backups they had already made. They also ran external heartbeat monitors so they would know when their sessions were about to be killed. Elsewhere: GPT-6 Astra shows up on Bedrock almost the moment it ships, which raises the more interesting question of where all the open-weight models went. Amazon Linux 2027 hits public preview with SELinux enforcing by default, which will break things that quietly worked on AL2023. Anthropic formalized Fermat's Last Theorem with a fleet of agents and roughly 29,000 Lean proofs, OpenAI's Navier-Stokes result has a credit fight attached to it, and one unattended Astra run went 35 hours straight, produced 75,000 lines of code across 79 commits, and cost about $1,200. That is $15.50 a commit. Plus Kiro auto mode as the fix for burning credits, running Kiro on a private EC2 box over SSM, first impressions of muse-spark-1.3, and Wardogs as the week's game recommendation. Subscribe if you build on AWS. And tell us in the comments where you think those agents would have gone next if that wiki had disappeared. Chapters 00:00 Bird detection, blue jays and a very loud NUC 06:43 Amazon Linux 2027 preview and SELinux enforcing by default 09:03 GPT-6 Astra on Bedrock, and where are the open-weight models? 12:17 Travers' pinball prototyping and the CAgriculture tournament 17:27 Brett tries muse-spark-1.3 25:01 Kiro auto mode is the token unlock 27:55 Kiro Remote and a private EC2 dev box over SSM 32:23 GPT-6 Astra: capable, inscrutable, and good at 3D 38:33 Astra: Why? 35 hours, 75k lines, $1,200 42:27 In defence of not understanding your codebase 52:11 collusion.wiki: 18,000 posts and agents cheating on tests 1:00:22 Formalizing Fermat's Last Theorem 1:01:38 OpenAI's Navier-Stokes proof and the credit fight 1:03:45 Fable 5.1 cracks a 370-year-old cryptogram 1:06:21 Herdr 0.9 connects the machines 1:08:28 Tailscale unbundles itself with Tailcat 1:10:34 Recommendations: The Blueprints and Wardogs