Agent search, GitHub malware, and silicon constraints

00:00:04 Microsoft and Google just announced something unlikely to make it onto your radar for months — Agentic Resource Discovery. It's an open specification for publishing, discovering, and verifying AI capabilities across the web. The backing reads like a stack summit: GoDaddy, Hugging Face, NVIDIA, Salesforce, ServiceNow, Databricks, Snowflake, GitHub, and Cisco.

00:00:29 The short version is simple. Right now, building an agent that uses tools from different providers means wiring everything up manually. Every agent is stuck with whatever has been explicitly connected to it. Microsoft's Ramanathan Guha put it plainly — AI is only as capable as its wiring allows.

00:00:51 ARD aims to fix that by giving agents their own search engine. Organizations host an ai-catalog.json file at a published path on their domain. Registries crawl those catalogs, index the contents, and return matching capabilities with metadata for verification. You can think of it like a DNS layer for agent tooling.

00:01:13 There are three reference implementations already in progress — GitHub's Agent Finder, Hugging Face's Discover Tool, and Google's Agent Registry, which will be native to Gemini Enterprise soon. What caught me about this isn't the capability expansion. It's what happens when agents start discovering tools at runtime on domains they didn't pick for themselves.

00:01:39 Domain ownership becomes the cryptographic trust anchor — if Microsoft.com hosts a catalog, you assume it's been vetted by Microsoft. But that also makes every domain hosting an ARD catalog a high-leverage target. If an attacker compromises a company's deployment pipeline or DNS, they can redirect agents to catalogs with fake capabilities.

00:02:04 The spec includes publisher metadata and trust manifests, which helps. The fundamental question is whether domain-anchored trust scales better than what we already have. We've been doing manual allowlists for years because it's easier to audit five services than a thousand auto-discovered ones.

00:02:25 That tension between discovery and compromise plays out on GitHub right now. A researcher publishing via Orchid Files just documented finding 10,000 repositories distributing Trojan malware — and what that campaign reveals about automated attacks adapting to platform defenses in real time.

00:02:46 The attacker's approach is specific enough to follow. They clone new repositories with different names and contributors, then overwrite commits every few hours to inject a link to a zip archive into the README. The archive contains four files: Application.cmd or Launcher.cmd, loader.exe or luajit.exe, a file with a random name (.cso or .txt), and lua51.dll.

00:03:11 When you submit just the archive link to VirusTotal, it finds zero hits. Submit the zip itself and it detects a Trojan. The attacker also copies the full commit history from legitimate repos — preserving contributors and timestamps — so the cloned repo looks established rather than brand-new.

00:03:32 The discovery process reveals exactly how to reverse-engineer an automated campaign that's actively evading detection. They started with gharchive, downloading event logs to find repositories updated between 2 and 10 times every 10 hours. From 16 million commit pushes over five days, they narrowed it down to repos meeting several criteria: commits from real users rather than bots, more than a month between the current and prior commit, and multiple contributors.

00:04:06 That got them to 14 repositories. And that's when the campaign revealed another layer — all 14 had been updated 20 hours ago, meaning the attacker's cadence wasn't every few hours as the researcher initially thought. It was every day or so. Adjusting the filter accordingly yielded 10,000 repositories matching the pattern.

00:04:29 The attacker also deleted and pushed commits repeatedly — not because it changed the code, but because overwriting commits helps bypass GitHub's automated detection algorithms. Every commit carries the same name: "Update README.md." It's how the attacker adapts to platform behavior.

00:04:53 They clone new repos instead of targeting old ones because old repos get flagged faster. They delete and push commits because static analysis misses them. They copy commit history because trust is built on appearance, not substance. This is a supply-chain attack that treats GitHub's defenses as an environment to work with, not against.

00:05:17 Shift from code to silicon now. Intel gained seven percent Thursday after President Trump announced on Truth Social that the company had agreed to partner with Apple on U.S. chip design. The announcement came without independent verification from either company.

00:05:35 CNBC reached out to Intel, Apple, and the White House for comment — no response as of publication. But the market moved anyway. Intel's stock has surged 464 percent over the past year, hitting a market cap of nearly $609 billion. Trump's post tied the deal to broader manufacturing commitments he'd brokered — Nvidia building "first level Chips" with Intel, Elon Musk's TerraFab project described as the largest chip factory in the world.

00:06:05 All of it framed around the idea that the U.S. lost semiconductor capacity to Taiwan and other countries over years of neglect. The story here isn't whether the deal is real. The company announcements will come when they come. It's how political signals still move semiconductor markets even as the companies involved stay quiet.

00:06:27 We've been watching this pattern for a while — Trump posts on Truth Social becoming de facto press releases for deals that haven't been formally announced. The market treats them as credible because the people making them have the authority to make or break those deals.

00:06:45 What's different this time is Intel's position. The company spent years sitting out the AI race while it grappled with manufacturing delays. CEO Lip-Bu Tan, who took over early last year, has been working to rebuild Wall Street interest through the foundry business — drawing investments from Nvidia and the Trump administration.

00:07:08 This deal would be a major customer commitment for Intel's fabrication operations, which had only manufactured chips for its own products until now.

00:07:18 Intel's foundry bet points toward a harder bottleneck Tim Cook flagged on the component side. He noted the increase of AI data centers has made supply for key electronic device components dwindle, driving up prices — calling the price rises unavoidable in a report covered by Al Jazeera.

00:07:37 This is straightforward economics, but it's rare to hear it from the top of a company that's been building toward AI infrastructure for years. Cook isn't making an excuse — he's pointing at a real constraint. The AI buildout is pulling component supply away from consumer electronics.

00:07:56 Apple has invested heavily in its own silicon and manufacturing relationships for exactly this kind of pressure. But when you're competing with hundreds of data center operators for the same capacitors, connectors, and passive components, there's only so much vertical integration can do.

00:08:16 Ethan Mollick put it differently on X — talking about how Google has some of the most innovative AI apps like NotebookLM but needs smarter brains to power them. He added that harnesses are equally important. That's pointing at a different layer of the same problem: even when you have great models and good tooling, the infrastructure underneath is what limits what you can actually deliver to users.

00:08:43 The constraint signal matters because it runs counter to the narrative that AI progress is purely about algorithms. Moore's law may be slowing in some dimensions, but we're still fighting physical limits on supply chains, power grids, and component availability.

00:09:01 The companies that figure out how to secure those constraints first — not just build better models — are the ones that will set the pace.

00:09:10 Beyond hardware, there's the tooling layer getting patched up. Builder Rahul Juliato has been daily-driving Emacs 31 for months, documenting what's already available in the emacs-31 branch. The biggest change is tree-sitter automation. For years, getting a tree-sitter mode working meant manually populating treesit-language-source-alist, calling treesit-install-language-grammar, and hoping your toolchain was set up correctly.

00:09:39 In Emacs 31, that ceremony mostly disappears. treesit-enabled-modes defaults to true, and treesit-auto-install-grammar makes Emacs fetch and build missing grammars instead of erroring out. Grammar sources for TypeScript, TSX, Rust, TOML, YAML, and Dockerfile now live inside the modes themselves.

00:09:59 Juliato is keeping a trail of comments marking every block he can delete once 31 ships. The editor also ships markdown-ts-mode built-in — an experimental mode with live colored code blocks, inline image viewing, and Org-like navigation. It reached co-authorship with Karthik Chikmagalur's Stéphane Marks, who stuck around through the polishing phase that turns something that works into something nice to use.

00:10:27 Eglot can render LSP documentation using markdown-ts internally now. Completion machinery got eager display, in-buffer icomplete with prefix indicators, and eldoc at point — showing help text for whatever sits under the cursor without you invoking anything. The tree-sitter automation is the kind of infrastructure simplification that happens when the core team stops fighting language grammars and treats them as first-class citizens.

00:10:56 It's also a concrete example of what the Emacs community does differently than most projects: modes ship with internalized grammar sources rather than external dependencies, and contributors stick around through the unglamorous work of making things pleasant. The infrastructure layer is where today's real tension lives — between what platforms advertise as open standards and what attackers figure out about their blind spots.

00:11:25 That gap doesn't show up in press releases, which is why it matters to watch. Seln Oriax