# AI Updates: OpenCode Key Breach, DeepSeek Flash, and NVIDIA Dory A breach at OpenCode's hosted Console shows that stored AI keys act like a spending card: if one leaks, someone else can run up your bill, so rotate it at… **Published:** 2026-10-19 · **Author:** Yash Thakker **Source:** https://explainx.ai/newsletter/ai-updates-opencode-key-breach-deepseek-flash-and-nvidia --- > **In short:** A breach at OpenCode's hosted Console shows that stored AI keys act like a spending card: if one leaks, someone else can run up your bill, so rotate it at the provider and cap spending. ## A Security Scare for AI Users ### OpenCode Console Breach: Stored Provider API Keys, 1.4% of Users, and What to Do Now OpenCode, the open-source coding assistant, says a breach on October 16 to 17 hit 1.4% of users of its hosted Console who had linked outside AI accounts. Stored keys, which let anyone spend your money at those providers, may be exposed. Anthropic says it revoked affected keys; OpenAI's refund plan is only reported. **What you can do:** If you linked an OpenAI, Anthropic or Google key in the OpenCode Console, replace that key in the provider's own dashboard, check recent usage, and set a spending limit. [Source](https://explainx.ai/blog/opencode-console-breach-stored-api-keys-distillation-attack-what-to-do-2026) ## AI News Briefs ### tv-vibehack: Debloat Your Android TV With Claude Code and ADB A new free GitHub project called tv-vibehack lets Claude Code, an AI coding assistant, switch off ads and tracking apps on an Android TV and install a simpler home screen. It disables apps rather than deleting them, and includes a restore script. Contributors have tested five TV models. [Source](https://explainx.ai/blog/tv-vibehack-debloat-android-tv-claude-code-adb-launcher-guide-2026) ### NVIDIA Dory 4B: A Looped Model You Can Dial Up at Inference NVIDIA released Dory, a small 4-billion-parameter AI model that reuses its own middle layers, and you choose how many times at run time. More repeats give better coding scores but cost more computing. NVIDIA reports 61.8 on a coding test at four repeats; no independent tests yet. [Source](https://explainx.ai/blog/nvidia-nemotron-labs-dory-4b-looped-model-nvfp4-adjustable-loops-swe-bench-2026) ### DeepSeek V4.1 Flash 1017: Open Weights, MIT License, Terminal-Bench 4.0 Up From 31.2 to 42.3 DeepSeek released an updated Flash model on October 17 with open weights under the MIT license, meaning anyone may use and modify it. Its own table shows one agent-coding score rising from 31.2 to 42.3. Those figures are self-reported, and no outside team has reproduced them yet. [Source](https://explainx.ai/blog/deepseek-v4-1-flash-1017-open-weights-mit-terminal-bench-4-42-benchmarks-pricing-2026) ## One Useful Idea ### What Is Gzip? Why Claude Code Now Responds 3x Sooner by Compressing Your Conversation Gzip is an everyday compression format that shrinks text without losing anything. Anthropic says Claude Code now compresses the conversation it sends each turn, and reports shorter waits before replies start. It has not said how much gzip alone contributes, since a separate faster cloud-startup change shipped too. **What you can do:** If you send large text to a service that accepts compressed uploads, test whether compressing first saves meaningful size; small files rarely gain much. [Source](https://explainx.ai/blog/what-is-gzip-claude-code-faster-first-token-compression-explained-2026) ## Upcoming Workshops ### AI Builder Workshop: Build and Ship with Claude Code ![AI Builder Workshop: Build and Ship with Claude Code](https://explainx.ai/workshops/covers/square/ai-builder.webp) *Live workshop · Nov 15–23, 2026* Four live Claude Code sessions take you from an idea to working automations, AI agents and a deployed app. It suits beginners through intermediate learners, and the workflow also carries over to Cursor. You practise building and shipping with guidance along the way. [Source](https://explainx.ai/workshops/ai-builder) ### Claude for Work: Become Your Team’s AI Person ![Claude for Work: Become Your Team’s AI Person](https://explainx.ai/workshops/covers/square/claude-for-work.webp) *Live workshop · Nov 28–29, 2026* Two live evenings for beginners who want to become their team's AI person. You build a knowledge base, an agent skill, a prompt library and your own Artifacts, and connect Claude to the apps you already use. [Source](https://explainx.ai/workshops/claude-for-work) ## Reader poll How much do you trust services that store your AI account keys? - Completely - Somewhat - Hardly at all - Not sure ## Sources - [OpenCode Console Breach: Stored Provider API Keys, 1.4% of Users, and What to Do Now](https://explainx.ai/blog/opencode-console-breach-stored-api-keys-distillation-attack-what-to-do-2026) - [tv-vibehack: Debloat Your Android TV With Claude Code and ADB](https://explainx.ai/blog/tv-vibehack-debloat-android-tv-claude-code-adb-launcher-guide-2026) - [NVIDIA Dory 4B: A Looped Model You Can Dial Up at Inference](https://explainx.ai/blog/nvidia-nemotron-labs-dory-4b-looped-model-nvfp4-adjustable-loops-swe-bench-2026) - [DeepSeek V4.1 Flash 1017: Open Weights, MIT License, Terminal-Bench 4.0 Up From 31.2 to 42.3](https://explainx.ai/blog/deepseek-v4-1-flash-1017-open-weights-mit-terminal-bench-4-42-benchmarks-pricing-2026) - [What Is Gzip? Why Claude Code Now Responds 3x Sooner by Compressing Your Conversation](https://explainx.ai/blog/what-is-gzip-claude-code-faster-first-token-compression-explained-2026) - [AI Builder Workshop: Build and Ship with Claude Code](https://explainx.ai/workshops/ai-builder) - [Claude for Work](https://explainx.ai/workshops/claude-for-work) --- From [explainx.ai](https://explainx.ai/newsletter). Read online, comment and subscribe: https://explainx.ai/newsletter/ai-updates-opencode-key-breach-deepseek-flash-and-nvidia