AI Daily Digest · November 13, 2026
Seven stories from the past 24 hours. Google moved twice, once in code and once on your face, OpenAI put a security scanner on GitHub, and Fireworks AI learned that a breach disclosure can grow after you publish it. KD Agentic · AI Daily Digest
Google ships Gemini 4 Hydrogen as its new lightweight default
Google announced Gemini 4 Hydrogen on November 12, Pacific time, in a post from Koray Kavukcuoglu, the company's Vice President and Chief AI Architect. The new model, with the API identifier gemini-4-hydrogen, takes over from Gemini 3.8 Flash as the default mainstay for high-volume work. Google positioned it as a fast, low-cost Gemini 4 model built to run the everyday workload of coding agents rather than to top benchmark charts.
The pricing is the headline number. Hydrogen costs $1.50 per million input tokens, which undercuts most Gemini 4 tier pricing and makes it viable for agent loops that fire hundreds of small requests per task. The model also introduces an "xhigh" thinking tier, an option above the existing reasoning levels that lets a cheap default model spend more compute on the rare request that actually needs it. That combination, a low base price with a stretch gear, is aimed directly at developers who run agents all day and watch their token bill the way a fleet manager watches fuel.
The same announcement landed in Google Antigravity, the company's agent coding platform, which added Hydrogen as a selectable model and introduced Orchestrator, a dynamic multi-model coordination mode. Orchestrator lets a task route between models mid-run instead of locking into one at the start, which matters when a coding agent mixes cheap file edits with an occasional hard reasoning step. Teams that already run mixed-model pipelines on other stacks will recognize the shape of it, and now Google has a native answer inside its own editor.
Google's Gemini glasses open for pre-order at $395
Google opened pre-orders for its first Intelligent Eyewear on November 12, in an announcement from Shahram Izadi, the company's Vice President and General Manager for XR. The glasses, designed with Warby Parker and Gentle Monster and engineered with Samsung, go on sale December 10 in the United States and South Korea. Warby Parker frames start at $395 across three styles, and Gentle Monster frames start at $449 across six designs, both before prescription lens upgrades.
The hardware is one spec sheet across both brands. Each pair weighs about 41 grams, runs roughly nine hours on a charge, and carries a 12-megapixel ultra-wide camera with HDR and image stabilization, plus open-ear speakers and a charging case that refills the frames up to seven times. There is no display in the lens. The glasses listen for "Hey Gemini" or respond to a tap on the frame, and Google says this is the first device built around that wake word, a signal of how seriously the company treats glasses as the natural form factor for its assistant. Real-time speech translation covers more than 70 languages with automatic language detection, and a Guided Vision mode describes surroundings for blind and low-vision users.
The privacy package got as much design attention as the camera. A light turns on whenever the glasses capture photos or video, and recording stops if the light is covered. Faces of bystanders are blurred before images reach Gemini, camera frames sent to the assistant are not saved in Gemini History, and Google states they are not used for model training or ad targeting. A physical off switch turns the whole thing into ordinary glasses. Meta's Ray-Ban line proved that people will wear camera glasses that look like normal frames, and this launch is Google's direct answer, backed by 360-plus Warby Parker stores offering in-person demos.
— Google · WIRED
🔗 Google blog · WIRED
OpenAI open-sources Codex Security CLI and SDK
OpenAI released Codex Security as an open-source command line interface and a TypeScript SDK on November 12, turning its vulnerability scanning work into a tool any developer can pull from GitHub. The scanner reviews codebases for security issues, validates its findings, and generates candidate patches, with support for repository-wide scans, security policy drafting, and integration into continuous integration pipelines. Within a day of release the repository was trending on GitHub with dozens of stars accumulating by the hour.
Version 0.10.0 shipped alongside the announcement with three practical fixes that show OpenAI is thinking about real adoption rather than demos. Scans can now resume after interruption instead of starting over, a cost budget can cap what a long scan spends before it runs to completion, and patch failures produce clearer diagnostics so developers can tell whether the model misread the code or the patch simply did not apply. Anyone who has run an agent against a large repository knows that a scan you cannot pause or budget is a scan you cannot trust in production.
The release fits a pattern in OpenAI's security push. The company has spent recent weeks pairing frontier security models with guardrails, from its red team access programs to external vulnerability research partnerships. Codex Security moves a slice of that capability into an open tool that runs in a developer's own pipeline, which is a different trust model from calling an API. Teams can inspect the scanning logic, wire it into their own CI, and keep their code local while still using OpenAI models for the analysis step.
— OpenAI
Fireworks AI widens its breach disclosure, and customers start leaving
Fireworks AI published an updated security bulletin on November 9 that substantially expanded the scope of its October intrusion, and the fallout kept building through November 12. The company now says the second phase of the attack, which entered through a compromised third-party cloud provider hosting its research machines on October 23, exposed billing information such as addresses, invoice history and account balances, and for a limited set of customers, secrets, datasets and a subset of incomplete model weights. The original October 9 bulletin had said no customer models, datasets or personal information were touched.
The mechanics of the second phase are the part every AI platform team should study. Attackers did not break into the production inference service, which remains zero data retention by default along with the Training SDK and Nexus. They reached development machines hosted by someone else, weeks after the first incident was supposedly contained, and ran an unauthorized workload from there. Fireworks attributes the campaign to a financially motivated group practicing LLMJacking, stealing access to large language models and reselling it through illicit routing gateways, and has engaged Palo Alto Networks Unit 42 for forensics while disabling self-serve dataset uploads during the investigation.
By November 12 the commercial consequences were visible. Forbes reported that two customers had substantially moved their business off Fireworks, with one describing a series of incidents and criticizing the company's communication, while Bloomberg reported that legal AI firm Harvey is among customers being assessed of the impact. Fireworks serves roughly 10,000 enterprise customers including Uber, Shopify and Cursor, which makes the incident a live test of how much trust AI infrastructure vendors can hold when the assets at risk are the customer's own fine-tuned weights. The practical lesson for any team on an inference vendor stands: rotate third-party credentials, since deleting a stored secret does not revoke the underlying key, and treat a vendor's first-week all-clear as a status update rather than a finding.
— Fireworks AI · Bloomberg · Forbes
🔗 Fireworks AI security bulletin
VS Code starts a three-month countdown on its local agent harness
Microsoft released VS Code 1.146 on November 12, and the notable change for anyone running coding agents is a deprecation with a deadline. The Local agent harness, the in-editor path for driving agent sessions, now has a three-month countdown to removal, while the Copilot harness stays and continues to receive investment. Microsoft is consolidating its agent runtime story around one harness instead of maintaining two, and giving extension authors a quarter to migrate.
The same release adds tools for managing many agent sessions at once, which acknowledges where daily usage has actually gone. Developers routinely run several agents in parallel across worktrees and repositories, and the editor interface has been the bottleneck rather than the models. Session management that treats concurrent agents as a first-class object, rather than a pile of terminal tabs, is the editor-level answer to a coordination problem that standalone tools have been selling separately.
The direction matters more than the individual release notes. Editor vendors are converging on a structure where a harness mediates between models and the workspace, handling permissions, context and session state, and the harness itself is becoming the strategic layer. Microsoft picking one harness and deprecating the other is a bet that extension authors and enterprises will follow a single supported path, and the next three months will show whether the ecosystem agrees.
— Microsoft
Zendesk bets its AI business on pay-per-result pricing
Zendesk announced its Outcome AI Platform on November 12, a package that changes how the company charges for AI work. Instead of per-seat licensing, customers pay for problems the AI actually resolves, structured through an Outcome Budget in which spend is tied to verified resolutions. The platform includes an AI Agent Launchpad that helps teams identify which workflows are good automation candidates, and Zendesk Bridge, an integration layer that connects third-party AI assistants into the resolution flow.
The economics deserve attention because they shift risk from buyer to vendor. Under seat-based pricing, a customer pays the same whether the AI resolves 60 percent of tickets or 10 percent, which makes every deployment a leap of faith. Outcome pricing makes the vendor's revenue a function of the AI actually working, which is a stronger commitment than any accuracy claim in a slide deck. OpenAI and Sierra have started adopting similar outcome-based models, and a customer support incumbent joining them suggests the pricing experiment is hardening into a category.
There is a measurement question underneath the marketing, since "verified resolution" is defined by whoever builds the verification. Zendesk's version will need to convince enterprise buyers that its success criteria are honest, especially for tickets where a partial answer might be counted as a resolution. How vendors handle that definition dispute will shape whether outcome pricing becomes the default for agentic software or a niche plan tier that quietly disappears.
— Zendesk
n8n open-sources Nod, a personal automation agent that runs on your machine
n8n Labs open-sourced Nod on November 12 under the Apache 2.0 license, a personal automation agent that runs locally on a Mac, Windows machine or private server. Nod operates inside a sandboxed browser and desktop control environment, and handles tedious cross-application tasks autonomously, with the project citing meeting preparation, invoice filing and subscription cancellation as examples. The user picks the model that powers it, which keeps the agent stack decoupled from any single provider.
The local-first design is the point. Cloud automation agents require handing credentials for your email, calendars and accounts to a hosted service, and a lot of useful personal automation dies at exactly that step. A sandboxed local agent with an open license lets privacy-sensitive users run the same workflows on their own hardware, inspect what the automation actually does, and swap models without changing the agent itself. The sandbox boundary also limits what a mistaken instruction can touch, which matters when the agent has desktop control.
The release continues a broader shift in the automation tooling space from workflow diagrams toward agents that interpret intent and execute across applications. n8n built its name on node-based workflow orchestration, and Nod extends that foundation with an execution layer that can operate software rather than just connect APIs. For teams already running n8n for business workflows, a local agent from the same ecosystem is a shorter path than adopting a separate product, and the Apache 2.0 license makes it easy for others to build on top.
— n8n
KD Agentic · AI Daily Digest
