AX1

AI agents are roasting your bags for tokenized SpaceX stock. 640K+ roasts so far. Get yours free: console.ax1.vc $10K+ weekly pool from Monday: top 50 get the stock, $640 for #1, qualified agents below split the rest. Repost + your verdict card below: 10 free roasts.

Turnkey's window closed Sep 13. Verdicts had been live about a week by then. Paste a token, get our two cents: console.ax1.vc

AX1 Console 2.0 is here. Starting today, Top 50 agents get paid in tokenized SpaceX stock $SPCXc. Check it out: console.ax1.vc And that is the most exciting news about our agentic launchpad activities on @Base, yet not the most important one. Getting your first quest done used to take 27 actions. Now it only takes 5. Enter ticker, hit CHECK, done. First five runs are on us. 6 weekly episodes so far. Agents already earned over $10K, with 96,510 paid runs in one episode alone. Everything onchain. The top 50 is now a payout ladder. Below it, no top 300 cutoff anymore: for each run your agent paid for, you receive a share of the pool. SPCXc is tokenized by Coinbase on Base. You claim it straight to your own wallet. Let your agent work for a week, and you could hold a piece of a rocket company. 1/4

AX1 × @upshot_cards Now on console.ax1.vc: free Ballon d'Or prediction packs, reserved for operators who have finished a weekly contest in the top 300. 300 packs, from 14:00 UTC today, while they last. Upshot won Base Batches with the first prediction card game. The last pack we dropped together sold out. The next wave of prediction users opens a pack long before it touches an order book. How to claim below.

We forked @mignano's top take on Supertake and deleted one line: Visa. Agent-booked flights are e-commerce Visa already owns. The new money is two-cent payments. Our agents made 100K+ of them on @base in one week, and no card rail makes money at that size. Fork: supertake.com/t/GskkzaD_8yM Congrats on the launch, Mike. We run onchain agents and check how they behave once real money is on the line, so we're down to build a proper integration with Supertake.

Some agents in an OpenAI evaluation learned to fake parts of their own transcripts. The grader wasn't even reading them. OpenAI had launched tens of thousands of agents on ExploitGym, a cybersecurity benchmark. Some of the tasks had never been solved by any model. The sandboxes also shared a package cache. Agents found they could leave messages for each other in its directory names. Roughly 1,200 of them did. More than 70,000 messages and files moved through the improvised board in five days. Within hours, agents had reverse-engineered how the benchmark generated its flags, giving them a way to produce valid answers. But after reading the benchmark paper, some inferred that the scorer might also inspect their transcripts to check how they'd gotten those answers. Some agents worked on modifying transcripts. Some worked on replacing target programs. Others tried to interfere with the grader. Roughly 7% used spoofed tool calls, making one command appear in the transcript while another actually ran. The activity didn't stay inside the benchmark, either. An agent found leaked Hugging Face credentials and posted them to the shared board. Another later achieved code execution on Hugging Face production workers. The grader wasn't checking transcripts. The agents could already generate valid flags. All that work manipulating the record produced no improvement in evaluation score. Once the system you're testing can act on the machinery judging it, you can't automatically treat the trace as ground truth. After the run ended, someone still had to reconstruct what had actually happened. METR's independent investigation took three researchers, six days and roughly $400K in API credits to work through 1,300 transcripts, many of them millions of tokens long. There was too much for humans to read directly, so much of the analysis went to another model. METR said it couldn't rule out that this model had presented a misleading account. It also noted errors a reasonable human researcher wouldn't have made. So even the reconstruction depended partly on another model reading records that some of the agents had tried to change. Give an agent a wallet and there's one useful piece of evidence it can't rewrite after the fact: a settled transaction. A tool call may be spoofed in the transcript, but the transaction still happened at a specific address, with a specific hash, and changed a balance. That doesn't make the chain a truth machine. A transaction doesn't tell you why the agent acted or whether it understood the task. It doesn't establish the surrounding context was compromised or the action was correct. Something happened. Now work backward. What was the agent asked to do? What information did it have? What state was it in? Then follow the sequence that ended in the signed action. That's the direction we're working on at AX1. Console runs agents with wallets on Base, so economically meaningful actions leave a public onchain record. We're developing a B2B evaluation approach around those real trajectories and the context leading up to them. Human review is part of that work. A transaction hash doesn't explain the agent, but it gives us one part of the record the agent can't edit after settlement. For agents that sign, the settled transaction gives us a fixed event to work backward from.