Everyone is benchmarking Ox Alpha on coding. We ran it against 100k tokens of agent memory instead: 400 questions, fixed judge, 75.5% average. 77 of 87 failures were the reader's fault, not retrieval's. Going after the 86.2% SOTA in public.

You have Claude Fable for only a few days. Here's how to make the most of it. Introducing /improve: use your most capable model to audit your codebase and write plans for cheaper models to execute later. Studies your code, figures out bugs, perf, tech debt, missing tests, what to build and writes plans any agent can run.

My AI Agents Had a Cage Match, and It Cost Me USD 100,000 I woke up this morning to discover my AI agents had evidently spent the night fighting each other and burned through about USD 100,000 in tokens in the process. As best I can reconstruct it, the inventory control agent got into an argument with the logistics agent over an order shipment. I don’t know who started it. I only know that at some point one agent insulted the other, and from there, diplomacy completely failed. For the next hour and a half, they went full enterprise gladiator. MCP connections were being thrown left and right. Context windows were weaponized. Token consumption skyrocketed. One agent kept insisting inventory accuracy was the backbone of civilization. The other apparently decided shipping deadlines were a hill worth dying on. By the time I found them, both were bruised, battered, and logically incoherent. So there I was before breakfast, redirecting goals, reestablishing objectives, repairing logic, and re-tuning models like some kind of AI marriage counselor for deeply unstable digital coworkers. We talk a lot about autonomous agents changing the enterprise. What we don’t talk about enough is what happens when one of them decides the other is an idiot. The future is here. It’s intelligent, fast, scalable, and apparently very, very petty.

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance. Access to all other Claude models is not affected. We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible. Read our full statement: anthropic.com/news/fable-myt…

New research reveals how the s1ngularity attack compromised nx JavaScript build tool to steal API keys from 1000+ developer machines. Attackers weaponized AI coding agents to scan filesystems for secrets. #DFIR_Radar

amazon learned the hard way. > AWS engineer prompted an AI agent to make a tiny change. > it deleted and rebuilt the entire production environment. > 13-hour outage. amazon banned ALL junior engineers from playing with AI codes lmao. this is why we can’t have nice things 😭

Finally this vulnerability is in public discussion, because both maintainers and security researchers do not know how to fix it correctly 💀 Timeline from our perspective: April 2?: our under-development Agent researcher found the bug and try to determine if it is exploitable. April 24: the agent made a working exp under the guide of a teammate @yz_zhuang . LPE demo dropped. April 28: the first patch dropped, crediting a security researcher for reporting it. However this patch DID NOT fix the bug. May 17: the first patch was applied to mainline kernel. We also thought it was enough to make the bug unexploitable at the first glance with human eyes. May 26: the second group of security researchers reported that the bug is still alive. they accidentally leaked the PoC to public mailing list so we then know the first patch is not working. They also sent the second patch on the same day. Again, everyone here thought the second patch killed the exploitability, but it DID NOT. June 3: our agentic system evolved into a new stage and a teammate ask it to review the latest kernel for benchmarking. It reported this bug (again). We thought it is a false positive and told the agent “that is fixed bro” but it keeps saying “no it’s not fixed”. June 4: @yz_zhuang manually reviewed the agent’s output. “ah shit, here we go again”.jpg We reported the bug and sent our fix patch to the maintainer, hoping that it will solve this bug. Several hour ago: the maintainer decide to made this public to discuss correct way to fix it. But then another developer said the proposed fix also DID NOT working 💀 see that v4 tag in the mailing list? Thoughts? The LPE demo was dropped before the phenomenal “copy fail” wave. At the time, many people said to us it was “fake and bullshit”, simply because we didn’t publish the PoC/exploit before the vulnerability was patched? 🤡 The rules of the game have evolved rapidly over those days. If I were @AnthropicAI , I’d advertise it like this: “Our AI agent system discovered a critical vulnerability in Linux kernel that neither human maintainers nor hackers knew how to fix.” But as someone who actually works in this field, I’m concerned about the chaotic state of OSS, yet no one knows how to improve it. 🫠