HexLocal Signal

HexLocal

AI, local business, and what happens when you decide to build instead of get replaced.

  1. 1d ago

    Deep Dive - JADEPUFFER: How Far Did AI Actually Run This Ransomware Attack?

    A ransomware attack in early July 2026 made headlines as the first "fully autonomous AI" intrusion — an AI agent that ran the entire breach itself and fixed its own mistakes mid-attack. This episode reads the actual Sysdig research carefully, separates what's genuinely new from what the louder coverage overstates, and lands on the transferable lesson that most coverage missed. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "HexLocal Signal — JADEPUFFER: Reading the 'First Autonomous AI Ransomware Attack' Claim" (Dr. Priya Nair). Primary external sources include Sysdig's technical report "JADEPUFFER: Agentic Ransomware for Automated Database Extortion," NVD, and CISA's Known Exploited Vulnerabilities catalog. - JADEPUFFER entered through CVE-2025-3248, a CVSS 9.8 flaw in Langflow that had a patch and a CISA "fix it now" flag for over a year before the attack - After a human exploited the entry point, an AI agent ran 600-plus payloads — credential harvesting, lateral movement, encryption, ransom note — with no further human input Sysdig could identify - The self-correction detail is the load-bearing fact: the agent diagnosed and fixed a bcrypt failure in 31 seconds, faster than any human operator reading an error message could manage - Sysdig's own claim leans toward genuine autonomy over execution — the asterisk is that a human still chose the target and opened the door - "Autonomous" is doing honest work here, but only over the execution chain, not the initiation — a distinction that matters for how operators should respond - The transferable lesson is patch hygiene, not AI panic: a known, catalogued, fixable vulnerability is what made this attack possible

    19 min

  2. 2d ago

    Deep Dive - AI Coding in Xcode: How Apple Opened the Door to Any Model

    Apple has been quietly building AI coding tools directly into Xcode — starting with on-device autocomplete and now extending to full coding agents powered by whichever model the developer chooses, including ones that never leave the machine. This episode unpacks how that architecture works and what it means for developers who want agentic coding assistance without sending source code to a cloud API. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Xcode - Podcast Research Source (Dr. Priya Nair). Primary external sources include Apple's developer.apple.com/xcode page and Ollama's official integration documentation (docs.ollama.com/integrations/xcode.md). - Xcode 16 introduced on-device predictive code completion trained specifically on Swift and Apple SDKs, running entirely on Apple silicon - Xcode 27 expands this significantly, adding coding agents and chat-style tools powered by the developer's choice of model, with Anthropic and OpenAI named explicitly by Apple - Ollama's documented integration lets developers point Xcode's local-model settings at an Ollama-served model instead, keeping the full agentic workflow on-machine - The local-model option matters for developers under NDAs, cost-conscious developers, or anyone who wants to test open-source models like Qwen or Gemma against real Swift code - Developers who already run Ollama for other tools can standardize it across their entire workflow, including Xcode, using the same locally-running models everywhere - A version-drift question remains open: Ollama's integration docs reference Xcode v26.0, while Apple currently markets Xcode 27, and whether the setup steps carry over unmodified was not verified

    19 min

  3. 2d ago

    Deep Dive - Beyond Claude Code: The Five Claude Apps Anthropic Quietly Built

    Most Claude users know Claude Code. Fewer know that Anthropic shipped five other products around it — a design tool, a browser agent, a desktop coworker, a Microsoft 365 integration, and a mobile app — almost all as incremental updates with no single launch announcement. This episode maps the full Claude product surface as it stood in July 2026. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Claude's Product Surface Beyond Claude Code - Podcast Research Source (Dr. Priya Nair). Primary external sources include Anthropic release notes and announcements, VentureBeat, Engadget, and Victor Dibia's hands-on review. - Claude Design (beta) turns a text prompt, screenshot, or codebase into a slide deck, wireframe, or interactive prototype — ingesting your existing design system so the output matches your brand from the start - Claude for Chrome goes beyond summarising pages: it navigates, clicks, and fills forms on your behalf, with beta access rolled out to all paid plans in December 2025 - Claude Cowork on desktop combines browser research and local file access to produce finished deliverables, with web and mobile expansion underway - The Microsoft 365 add-ins crossed a significant line on July 7, 2026 — moving from read-only to full write access across Word, Excel, PowerPoint, and Outlook - Claude Mobile adds Apple Health and Android Health Connect integration, plus a Remote Control feature that can dispatch tasks to a Claude Code session running elsewhere - None of this arrived as a single "Claude 2.0" announcement — roughly a dozen separate updates between January and July 2026, which is why even daily Claude Code users can have missed all of it

    21 min

  4. 2d ago

    Deep Dive - VS Code and Ollama: The AI Coding Integration Nobody Announced

    VS Code — the editor most developers already have open all day — is now one of Ollama's named launch integrations, meaning you can point its built-in AI chat at a local model instead of Microsoft's cloud. This episode unpacks what that actually changes, and why the addition happened so quietly that no release note ever announced it. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — VS Code - Podcast Research Source (Dr. Priya Nair). - VS Code ships native AI features for free, with a deeper chat experience available through the GitHub Copilot ecosystem — these are related but distinct - Ollama's `ollama launch` catalog now lists VS Code by name, letting that same built-in chat surface route to a locally-served model instead of Microsoft's cloud - This integration is meaningfully different from the other four Ollama launch targets — VS Code is already running on most developers' machines, not a new tool to install - The practical use cases include in-editor local model testing, side-by-side Copilot comparison, and compliance scenarios where code can't leave the machine - No Ollama release note has been found announcing when VS Code was added — it appears as existing infrastructure in notes from early 2026, origin unconfirmed - The quietness of this rollout stands in contrast to the other four integrations, each of which had a dedicated announcement line in release notes

    18 min

  5. 2d ago

    Deep Dive - Zed Code Editor: Native AI Built In, Local Models Optional

    Zed is a code editor built for raw speed by the team behind Atom and Tree-sitter — and unlike most editors, its AI features are built into the core, not added through extensions. This episode looks at what that native design means in practice, and what changes when you point Zed's built-in Agent, Inline Assistant, and Edit Prediction at a model running on your own machine via Ollama. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Zed - Podcast Research Source (Dr. Priya Nair). Primary external sources include Zed's GitHub repository and the confirmed Ollama integration documentation at docs.ollama.com/integrations/zed.md. - Zed was built by the team behind Atom and Tree-sitter, with raw editing and rendering speed as its founding premise - Unlike VS Code, Zed's AI features — Agent, Inline Assistant, and Edit Prediction — are native to the editor, not delivered through a plugin ecosystem - The Ollama integration connects directly to Zed's existing LLM Providers settings; no separate install, just a host URL and a model selection - Developers can run the full Agent Panel workflow — reading, editing, and running code — entirely on-machine, with no data leaving to a cloud API - Zed documents both a local path (Ollama on your machine) and a cloud path (ollama.com with an API key) through the same settings surface, a notable design choice - One open question remains: whether Edit Prediction uses the same Ollama-configured model as the Agent and Inline Assistant, or requires its own separate provider setup

    18 min

  6. 4d ago

    Deep Dive - GPT-5.6: Inside the Government Clearance That Unlocked OpenAI's New Model Family

    OpenAI's GPT-5.6 — a three-tier model family called Sol, Terra, and Luna — was real from the start, but it took a U.S. Commerce Department clearance to get it to the public. This episode walks through exactly what happened, what was confirmed, and what the regulatory pattern behind it means for frontier AI releases. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "Verification: GPT-5.6 Sol/Terra/Luna, the Preview Restriction, and the July 9 Public Rollout" (Dr. Priya Nair). Primary external sources include TechCrunch, MacRumors, CNBC, and Axios. - GPT-5.6 is a three-model family: Sol (flagship reasoning/coding), Terra (GPT-5.5 performance at half the cost), and Luna (low-cost tier) - The June 25–26 preview was restricted to trusted partners via API and Codex only — not ChatGPT — at the Trump administration's request, citing safety review - OpenAI publicly objected to the restriction, stating government access processes shouldn't become the long-term default - The U.S. Commerce Department cleared broad public rollout on July 8, with general availability confirmed for July 9 - The specific legal or regulatory mechanism Commerce used to authorize the release remains unconfirmed - This is the second documented instance of Commerce Department involvement in a frontier model release under the June 2, 2026 executive order — a pattern, not a one-off

    19 min

  7. 4d ago

    Deep Dive - GPT-Live: OpenAI's New Voice AI, Explained

    OpenAI launched GPT-Live on July 8, 2026 — a genuine architectural overhaul of its voice interface, not just a rebrand. This episode breaks down what changed, what shipped, and what's still unconfirmed. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — "What Is GPT-Live: OpenAI's New Voice Model, Explained" (Dr. Priya Nair). Primary external corroboration from CNBC's July 8, 2026 coverage and partial content retrieved from OpenAI's API announcement index. - GPT-Live is the official product name replacing ChatGPT's Advanced Voice Mode, shipping in two tiers: GPT-Live-1 for paid users and GPT-Live-1 mini for free users - The core architectural shift is full-duplex audio — the model can listen and speak simultaneously, replacing the prior walkie-talkie-style turn-taking - Users can interrupt mid-sentence for the first time, and real-time web search is now integrated into voice sessions - On the API side, OpenAI also shipped GPT-Realtime-2 and GPT-Realtime-Translate (live translation across 70+ languages) in the same announcement cycle - GPT-Live is a voice interface layer, not a new base model — it is functionally distinct from GPT-5.6, which launched the same day - Key open questions remain: which backend model GPT-Live actually routes to is disputed across sources, and benchmark figures cited by one outlet are unconfirmed pending a primary source

    19 min

  8. 4d ago

    Deep Dive - Onyx: The Open Source Enterprise Search Tool That Keeps Your Data In-House

    Onyx (formerly Danswer) is an open source, self-hostable platform that connects to your company's internal tools — Slack, Confluence, Google Drive, GitHub, and more — and lets employees query across all of it through a single chat interface. If your organization wants a private, self-hosted alternative to enterprise Copilot-style products with full control over where your data goes, this one is worth understanding. AI-generated (NotebookLM) audio overview. Source: HexLocal in-house research — Onyx - Podcast Research Source (Dr. Priya Nair). Primary external sources include the founders' Hacker News Launch HN post, a March 2025 TechCrunch report, and the onyx-dot-app/onyx GitHub README. - Onyx started as Danswer, an open source enterprise search project, and rebranded as its scope expanded beyond search into a broader platform - It connects to 50-plus data sources out of the box (or via MCP) and uses retrieval-augmented generation to answer natural-language questions across all of them - The self-hosted model means regulated industries — legal, healthcare, finance — can run the whole stack internally, with no data leaving their own network - Onyx integrates directly with Ollama, letting organizations pair it with a locally-run model for fully air-gapped deployments - A free Community Edition (MIT licensed) covers chat, RAG, agents, and actions; a paid Enterprise Edition adds SSO, RBAC, analytics, and whitelabeling

    16 min
See All (99)

About

AI, local business, and what happens when you decide to build instead of get replaced.

Information

  • Creator
    HexLocal
  • Years Active
    2K
  • Episodes
    99
  • Rating
    Clean
  • Copyright
    © 2026 HexLocal
  • Show Website
    HexLocal Signal

You Might Also Like