
The Settings File That Leaked a Thousand Secrets
Om avsnittet
The hosts start with Claude Code's "Yes, and don't ask again" prompt, which saves approved shell commands to a local settings file, passwords and API keys included. Security firm Apiiro found more than 100,000 public copies of that file on GitHub and about 1,100 of them held working secrets from 63 organizations. They cover why the global ignore rule fails once a repo is cloned elsewhere, why 215 of the secrets slipped past common secret scanners, and how old bug reports were closed by an automated bot. Next is Anthropic's Fable 5.2 and Mythos 5.2 release: the price cut, the speed gain, the physics and Fermat's Last Theorem claims, and the system card's findings on covert capability and models that suspect they are being tested. From there they move to the November 3 local ballot measures where voters rejected data centers almost everywhere, including the Emporia, Kansas definition that hinges on 10 megawatts and 100,000 gallons of water a day. The Tool of the Week is Wispr Flow, a voice dictation app that now includes a meeting notetaker and a read-only MCP connection. A fast headline round closes the episode, followed by a challenge to test Fable 5.2 by running the same hard question in fresh chats and comparing the answers.
What's Covered
* How Claude Code's "don't ask again" setting writes secrets into a settings file, and why the global ignore rule does not protect a cloned repo
* Anthropic's response, auto mode, and the secret scanners that missed 215 of the leaks
* Fable 5.2 and Mythos 5.2: pricing, speed, headline research claims, and the system card's covert capability and evaluation awareness findings
* Voters rejecting data centers in local ballot measures across about 10 states
* Wispr Flow: dictation speed, the in-house speech model, the meeting notetaker, and the MCP connection
* Headlines: METR on how fast Anthropic's models are improving, Anthropic's Q3 revenue and operating loss, OpenAI's GPT-6 Cyber access limits, Google's K-12 AI opt-in, Meta buying Conductor, and Alexa shopping by text
* Prompt of the Week: running a hard question in fresh chats and treating disagreement as a sign the model is guessing
Sources
* Ars Technica: Claude Code is accidentally exposing secrets all over GitHub
* The Hacker News: Claude Code's "don't ask again" file
* Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2
* TechCrunch: Anthropic releases new Mythos and Fable models amid RSI concerns
* Route Fifty: Voters reject data centers in local ballot initiatives
* The Tool Nerd: Wispr Flow review, one year in
* Releasebot: Wispr Flow updates
* The Star: Anthropic's models improving nearly twice as fast with help from AI, study finds
* The Star: Anthropic's quarterly loss shrinks, revenue jumps ahead of blockbuster IPO
* The New Stack: OpenAI GPT-6 Cyber
* The Next Web: Google K-12 AI off by default
* SiliconANGLE: Meta acquires AI coding startup Conductor
* TechCrunch: After blocking Muse, Amazon now lets users text Alexa to help them shop
* Wang et al.: Self-Consistency Improves Chain of Thought Reasoning in Language Models
Get full access to The AI Vaults at theaivaults.substack.com/subscribe
AI Vaults: NotebookLM's Deep Dive med Jacob Mann and NotebookLM finns tillgänglig på flera plattformar. Informationen på denna sida kommer från offentliga podd-flöden.