Today’s lead · OpenAI News
OpenAI ships GPT-6 Cyber into the vulnerability-hunting arms race
OpenAI’s cyber-specialized model moves frontier capability into authorized vulnerability discovery and patching, with early use at Mozilla and others. The practical upside is faster defensive audit work; the uncomfortable bit is the same capability curve also lowers the cost of finding subtle bugs.
Top signals
5 more
Gemini Omni 1.2 Flash makes video generation longer, cheaper, and more editable
Google’s updated video model now supports native 15-second generation and editing, multi-turn conversational changes, and better reference adherence. For builders, the useful change is not just clip length; it is cheaper iteration at $0.07 per second for 720p, with production availability across Google’s API and cloud stack.
NVIDIA opens Cosmos 3 checkpoints for interactive simulation work
NVIDIA added four open-weight Cosmos 3 Nano checkpoints for driving and robotics simulation. They are useful building blocks for synthetic data, streaming sensor rollouts, simulator-to-RGB rendering, and bimanual robot action prediction, but the post is clear that measuring policy-learning fidelity remains unfinished work.
Axiom Math formally verifies the Stacks Project with AI agents
Axiom Math says its agents translated the 7,654-page Stacks Project into roughly 16 million lines of Lean. The important signal is not AI “doing math” in prose, but AI helping produce machine-checkable foundations that researchers can actually build on.
SenseNova 6.8 Code2Video points video generation back toward editable source
SenseTime’s Code2Video release is interesting because it makes video artifacts inspectable: prompts become executable code, not only sampled pixels. That trades photorealism ambition for editability, exact text, timing control, and composable workflows that can call other generation or creative tools.
Cursor starts closing the localhost-to-live-app gap
Cursor is adding live app building and link sharing inside the editor. The pitch is pragmatic: agents can already write code, but deployment, databases, and security checks are where prototypes stall. Treat this as a workflow claim until builders test the guardrails on real apps.
Tools & repos
4 selected
Glasshand
Glasshand rents real iPhones and Macs to agents like Claude Code or Codex, giving them dedicated hardware for tasks such as app testing, posting, or booking flows, then factory-resets the device afterward.
Rerun
Rerun turns existing ChatGPT or Claude subscriptions into scheduled or trigger-based AI coworkers with tools, a computer, and a shared workspace. The notable bit: it is now AGPL v3 and self-hostable.
Pushary for Teams
Pushary routes AI-agent permission prompts to the right teammate via Mac, phone, or dashboard, then keeps an audit trail of who approved what. That is useful plumbing for teams letting agents touch real systems.
Clocknext
Clocknext is billing infrastructure for AI-native apps and agents, focused on turning real-time AI usage into billable metrics while tracking costs, margins, and profitability across pricing plans.
Blogs worth your time
4 reads
PDD Kit tries to make agent-written code answer to proofs, not vibes
Microsoft Research’s PDD Kit turns developer intent into Verus specs, proves Rust code against them, and tests trusted boundaries. It is still a research prototype, but the framing is right: review intent, not every generated line.
LangChain explains why serious agents need async subagents
LangChain’s Deep Agents now support async subagents: an orchestrator stays responsive while background workers handle long tasks. This is becoming core agent infrastructure, especially when approvals, notifications, and parallel work all matter.
NVIDIA DSX OS is about keeping GPU clusters production-ready, not just installed
NVIDIA’s DSX OS stack packages cluster baseline recipes, workload certification, and continuous fault handling for Kubernetes GPU fleets. The builder takeaway: health checks are not enough when one slow GPU can poison distributed training goodput.
NVIDIA’s memory story for agentic inference is full-stack co-design
This NVIDIA deep dive argues that HBM capacity alone is the wrong metric for agent workloads. The levers are NVFP4, expert parallelism, Dynamo KV reuse, Vera CPU memory, and CMX storage.
Community discussions
2 threadsA voice-agent failure that every real-world automation builder should steal as a test
The agent correctly treated rat poison as an emergency, then falsely promised the vet was available. The thread’s useful lesson: voice agents need hard boundaries, live-state checks, and adversarial tests from domain experts, not just happy-path demos.
AI site builders are reopening an old SEO wound: rendered JavaScript
The post warns that AI website builders using client-rendered React can hide content from crawlers. Whether every claim generalizes or not, the builder check is simple: view source and ensure core content exists as HTML.
Funding & acquisitions
2 moves
Dropbox buys Tigris Data for $350M to build AI-agent storage infrastructure
Dropbox is acquiring Tigris Data to support DBX Agent Cloud, an S3-compatible storage platform for AI models and agents. This is Dropbox repositioning storage around agent workflows, not just file sync.
Venomics AI raises $2.7M to mine venom biodiversity for drug discovery
IISc spin-out Venomics AI raised a pre-seed round to expand Venom Atlas, its molecular dataset of venom-derived compounds, and strengthen AI plus lab screening for therapeutic leads.
Bengaluru radar
2 events
The Agentic Room
One-day agent buildathon where your AI agent is handed to a stranger for a real task while judges watch. Useful pressure test for demo-first builders.
Design School S2: AI took my brief, not my job
AI Learn Circle and Paytm host a practical design session on briefing Claude, iterating outputs, and using AI in B2B product design workflows.