The Signal — June 3, 2026
Cybersecurity gets frontier-model reinforcements, knowledge work gets a new platform play, and open-weight computer agents land on local hardware for the first time.
Anthropic Expands Project Glasswing to 200 Partners
Anthropic is quadrupling the reach of Project Glasswing, its restricted cybersecurity program built around Claude Mythos Preview. The initiative is growing from roughly 50 organizations to approximately 200, adding around 150 new partners in critical infrastructure (power grids, water systems, healthcare networks, and communications providers) across more than 15 countries.
In two months, existing partners including Cloudflare have used the model to surface over 10,000 high- or critical-severity vulnerabilities. That volume reflects both the scale of unpatched exposure in critical infrastructure and the capability gap that frontier models can fill when pointed at defensive work.
Anthropic paired the expansion with a pointed warning: comparable models from other labs could arrive within six to twelve months, potentially without the access controls Glasswing enforces. Anthropic is trying to build the defensive playbook before the offensive one writes itself, and the open question is whether other labs will adopt similar restraints or skip them entirely.
Sources: Anthropic · TechCrunch · CyberScoop · Yahoo Finance · Engadget
OpenAI Expands Codex Beyond Coding Into General Knowledge Work
OpenAI is repositioning Codex from a developer tool into a general-purpose productivity engine. The updated product now targets research, analysis, knowledge management, and workflow automation, work that every role in an organization touches, not just engineers writing code.
The move is accompanied by a full report on what OpenAI calls "the next era of knowledge work," a framing that signals serious enterprise ambitions. Where Codex once competed primarily with GitHub Copilot and similar code-completion tools, it now stakes a claim against the broader productivity suite market, territory occupied by Microsoft, Google, and a growing crowd of AI-native startups.
The strategic logic is straightforward: coding assistants have proven the model, and the same retrieval, reasoning, generation, and execution capabilities generalize to structured knowledge tasks. The question is execution. Enterprise buyers will want to see that a tool born in code editors can handle the messy, unstructured reality of business workflows without hallucinating its way through compliance documents or client reports.
Sources: OpenAI · OpenAI Report (PDF) · The Deep View
H Company Releases Holo3.1: Computer Use Agents Go Local
H Company released Holo3.1, a family of four open-weight models ranging from 0.8 billion to 35 billion parameters, all built for one job: letting autonomous agents control computers without cloud dependencies. The models come in quantized formats (FP8, Q4 GGUF, NVFP4) ready for local deployment on consumer hardware, and for the first time, mobile devices.
On the OSWorld benchmark, which measures real-world computer operation, Holo3.1 matches or exceeds proprietary alternatives that cost orders of magnitude more to run per query. The 0.8B model is small enough for a phone. The 35B model uses a mixture-of-experts architecture (35B total, 3B active) that keeps inference fast on a single GPU. Both extremes represent the same bet: computer-use agents only become broadly useful when running them does not require a credit card attached to an API.
This release matters because it shifts the agentic AI conversation away from which cloud provider you pick and toward what your own hardware can do. When a sub-billion-parameter model can navigate desktop applications competently, the barrier to deploying personal AI assistants drops from "enterprise budget" to "decent laptop." The models are Apache 2.0 licensed and available on Hugging Face.
Sources: Hugging Face Blog · AI Weekly · KeepingUpWith.ai
On the Editor's Desk
The largest held story was the agentic hardware-software stack unveiled at a major developer conference yesterday. We covered that partnership extensively in our June 1 and June 2 editions; today's details are incremental extensions rather than a fresh development. The EU's approaching enforcement deadline for its regulation on general-purpose models (August 2) is important context but not a breaking event yet. An insurance company deploying automated claims processing is a solid case study but stays a single-deployment story. A youth safety policy announcement reads as positioning rather than a capability shift. All four remain on the watch list.
