Today’s lead · x.com
OpenAI puts Codex on vulnerability duty
OpenAI expanded Daybreak with a Codex Security plugin for finding, validating, and fixing vulnerabilities, the full GPT-5.5-Cyber model for defenders, a partner program, and Patch the Planet with Trail of Bits reviewing findings before maintainers see them. The useful design choice is that AI-generated bug reports get a human security-engineering triage layer, rather than dumping more noisy issues onto open-source projects. This caught my attention because security is becoming one of the first places code agents need a credible verification loop, not just a faster autocomplete.
Top signals
8 more
Reflection buys a frontier-scale open-model runway
Reflection AI signed a SpaceX compute deal worth up to $6.3B, paying $150M per month from July 2026 through 2029 for GB300 capacity at the Colossus 2 data center. For an open-weight lab founded by former DeepMind researchers, that is a serious statement that open models now need the same industrial-scale training budgets as closed labs.
Dify’s agent cloud shows why tenant isolation matters
Researchers disclosed DifyTap, four flaws in Dify that could expose AI chats, uploaded files, and tracing data across tenants in an open-source agentic workflow platform with more than 146,000 GitHub stars. The uncomfortable lesson is that agent platforms are now multi-tenant data planes, so boring authorization checks are just as important as prompt-injection defenses.
AgentSpace makes agents look like managed coworkers
AgentSpace was open-sourced as a shared workspace where teams can assign agents roles, owners, schedules, permissions, approvals, and audit trails, with AgentRouter sending work across Claude Code, Codex, OpenClaw, Hermes, nanobot, and other runtimes. I like the framing because the hard part of enterprise agents is not spawning another bot; it is making identity, accountability, and execution visible to the whole team.
Grok Build adds a long-running goal mode
xAI introduced /goal in Grok Build, a mode for autonomous tasks that runs multiple rounds of subagents to implement and verify one objective. The interesting bit is the product shape: coding tools are converging on planner-builder-verifier loops as a first-class command, not a pattern users have to assemble by hand.
Claude Code’s team is measuring the messy parts
A recap from Claude Code and Cowork lead Fiona Fung says Anthropic engineers are shipping about 8x more code than a year ago, which has made verification the bottleneck; the team tracks “bad” unrecoverable failures, “sad” recoverable UX pain, and even how often users swear at Claude Code. That is the kind of production signal I care about because evals for coding agents have to include user frustration, not just benchmark pass rates.
Video editing models get their own arena
Artificial Analysis launched a Video Editing Arena covering Seedance 2.0, Runway Aleph 2.0, Wan 2.7, HappyHorse 1.0, Kling 3.0 Omni, and SkyReels V4, with voting live and a leaderboard planned. Treating text-directed video editing as a separate eval category is the right move, because instruction-following over time, audio, objects, physics, and effects is a different problem than raw generation quality.
Groq refills the tank after Nvidia’s IP move
Groq confirmed a $650M round after Nvidia licensed its LPU technology and hired away its founder, with the company now emphasizing a neocloud inference business spanning 13 data centers, five million developers, and trillions of tokens per week. The question is whether Groq can stay differentiated when its hardware IP is now also in Nvidia’s hands, but inference demand is big enough that the pivot still matters.
Nvidia’s cooling story moves to 45°C loops
Nvidia says its Rubin-era DSX AI factory design uses 100% liquid cooling with 45°C coolant, enabling dry-cooler operation in favorable climates and cutting facility cooling water from roughly 2.6M gallons per MW per year to near zero. TechCrunch correctly points out that this does not solve upstream power-generation water use, but at the rack and facility layer it is still a major infra design shift for dense AI clusters.
Tools & repos
3 selectedcalesthio/OpenMontage
World's first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turn your AI coding assistant into a full video production studio.
palmier-io/palmier-pro
macOS video editor built for AI
jamiepine/voicebox
The open-source AI voice studio. Clone, dictate, create.
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Funding & acquisitions
5 movesSkybridge
The full-stack open source React framework for MCP Apps
AgentX
Evaluate AI agent, pinpoint issues, and fix with one click.
Alai 2.0
AI design partner for presentations, social posts, and more
HAQQ Legal AI on Mobile
Bringing legal understanding to anyone with a phone
readywhen
Your 24/7 AI Chief of Staff for commitments and follow-ups
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.