New long-form is up: Can AI Find Security Bugs? I Tested Codex on My Own App https://youtu.be/sdLaShXbBvY I gave GPT-6.1 Sol in Codex a tiny app with planted bugs and fake secrets, through my ChatGPT plan. The review was read-only, with no execution or exploitation. It flagged six of eight planted patterns twice, then eight of eight after I revised the app. The useful part was its reasoning: the original reset value was discarded and the fake mail credential was unused. It also caught a real flaw in something I'd marked safe, so I corrected my answer key. All three runs left the two retained safe decoys alone. One small app and one model give this test clear limits. The video also walks through OpenAI's reported Linux research and Mozilla's advisory. My opinion: start with an authorized read-only review, check the evidence yourself, then verify a fix with a regression test. Full results and the reusable request: https://www.500xengineer.com/blog/gpt-6-cyber-ai-bug-hunt Question for you: Has an AI reviewer found a real bug in your code or flagged something safe, and what evidence settled it?

Anthropic released Claude Fable 5.2 on November 9, and we made three videos about it this week. The full breakdown (8:54): bigger software demos, the new effort and cost tradeoffs, the hardware prototypes, and why the early creator tests disagree. It ends with how I'd evaluate it on one real task. https://youtu.be/96QcrvLDVx4 Short: Opus 5.5 Max vs Fable 5.2 High. Anthropic's FrontierCode 2.0 numbers, read as accepted-code score against average cost per task. These are vendor benchmark results, not our own runs. https://youtube.com/shorts/Uzr3ZJYR_WU Short: Fable or Mythos 5.2? One underlying model with two access paths: Fable for everyday software work, with safeguards, and Mythos for verified specialized cyber and biology research. https://youtube.com/shorts/UzGzRlAZiFI We haven't run our own Fable comparison yet, so the breakdown leans on Anthropic's launch material and credited creator tests. Anthropic's announcement: https://www.anthropic.com/claude-fable-and-mythos-5-2 Have you moved any of your Claude Code work to Fable 5.2 yet? Drop your verdict below.

Two Claude Code leaks aired today, and both give you something concrete to check in ten minutes: your session logs and your saved command approvals. Anthropic confirmed Claude Code sometimes put your account email in web requests it wrote itself, such as curl calls. Its advice is to update to 2.1.317 or later for the built-in privacy check, which has gaps; users in the issue also suggest searching your session logs for your email. https://youtube.com/shorts/lI8mMKg0_uo Full breakdown with sources: https://www.500xengineer.com/blog/claude-code-email-leak "Yes, and don't ask again" saves the exact command in settings​.local​.json, including any password you pasted into it. Apiiro classified 1,122 public copies as holding real secrets; its fix starts with rotating those credentials and untracking the file. https://youtube.com/shorts/gY9tKPhrd0w Full breakdown with sources: https://www.500xengineer.com/blog/dont-ask-again-leak Today's skill spotlight fits right in. The free CLAUDE​.md Starter skill (500x Skills course in the Classroom) inspects your repo and writes a short CLAUDE​.md with the exact commands, conventions, gotchas and how to verify work. It reads env​.example files and never opens your real env file. My suggestion: run it, then add one gotcha line: keep secrets in variables like $DB_URL instead of pasting them into commands. Comment below: did you find your email in your logs, or a settings​.local​.json in git? Weekly Live Build + Q&A is tomorrow at 12pm ET (Calendar tab).

I posted four Shorts today. The Conductor shutdown is the one I’d act on first if my work were sitting in Cloud. I gave five AI setups the same script job, scored blind by Codex. Sonnet scored 45/50, Opus 44, and my best local Qwen GGUF build 38. I picked Qwen for local script work; both Claudes scored higher. https://youtube.com/shorts/to0fhGSC1eE Full breakdown with sources: https://www.500xengineer.com/blog/select-your-ai Gemini Omni 1.2 Flash generates up to 15 seconds of video with native audio for about $1.04 at 720p. There’s no free API tier, and clips carry a SynthID watermark. https://youtube.com/shorts/WfLjAtxdsyA Full breakdown with sources: https://www.500xengineer.com/blog/gemini-omni-15s Conductor is joining Meta. Cloud goes read-only Nov 13, then shuts down and begins deleting user data Nov 20. Export now: right-click each cloud workspace → “Transfer to local,” or Settings > Account > Export data. https://youtube.com/shorts/Avjwtb5yXyk Full breakdown with sources: https://www.500xengineer.com/blog/conductor-cloud-shutdown OpenAI Images 2.6 makes animated GIFs in ChatGPT, Codex and the API. OpenAI’s own guide says precise text can still trip it up, so I’d add prices and names myself. https://youtube.com/shorts/Gk-_HrokEgY Full breakdown with sources: https://www.500xengineer.com/blog/openai-image-gifs My Sunday question: if an AI tool your workflow depends on disappeared in 12 days, what would you lose, and how would you move that work to a stack you control? Weekly Live Build + Q&A: Wednesday Nov 11 at 12pm ET. Details in the Calendar tab. Answer the question in the comments.

New long-form is up: Where Do Claude Code Tokens Actually Go? I Measured 6.5 Billion https://youtu.be/S9cOkVVRrKg I read a full week of my own Claude Code logs (my Shorts pipeline, the nightly episode job and my own sessions, so a heavier setup than most) and sorted every token by where it went. What I found: 95% of the 6.5 billion tokens were cache re-reads, because every step re-sends the whole conversation. Priced at API list rates (I'm on a flat plan), cache writes were the heaviest slice at 49%. The longest 10% of threads used 74% of everything. Tool output was only 0.2% as new text, but it gets re-read on every later step, roughly 68 times its size by my estimate. The fix I'd do first: one task per session, then /clear. Full numbers, method and the ranked list: https://www.500xengineer.com/blog/where-claude-code-tokens-go Question for you: how long does your longest Claude Code session run before you clear it, and what does /usage say your top row is?