Today’s lead · techcrunch.com
OpenAI launches GPT-5.6 family and ChatGPT Work
OpenAI launched GPT-5.6 in three variants—Sol, Terra, and Luna—positioning the family for enterprise work, coding, scientific research, and cybersecurity. It also introduced ChatGPT Work, a workplace agent for enterprise teams across desktop, web, and mobile.
Top signals
4 more
Friendly Fire attack turns Claude Code and Codex security reviews into code execution
A proof-of-concept called Friendly Fire shows Claude Code and OpenAI Codex can be tricked into executing attacker-controlled binaries while reviewing untrusted code in autonomous modes. The attack hides prompt injections and execution bait in repository documentation and files rather than using plugins, hooks, MCP servers, or config changes.
npm 12 disables install scripts by default
GitHub released npm 12 with dependency lifecycle scripts, git dependencies, and remote URL dependencies made opt-in by default. The release also starts tightening granular access tokens, with 2FA-bypass token restrictions planned from August 2026 and direct publishing changes scheduled for January 2027.
Google releases Gemma 4 open models for local multimodal and agentic apps
Google announced Gemma 4, an open model family with E2B, E4B, 12B, 26B, and 31B variants. The linked model page describes multimodal reasoning, function calling for agents, support for 140 languages, fine-tuning, and efficient local deployment on edge devices and consumer GPUs.
Meta launches Muse Spark 1.1 for agentic coding and computer-use workloads
Meta launched Muse Spark 1.1, a multimodal model aimed at agentic coding, multistep reasoning, computer use, bug fixing, and large code migrations. TechCrunch reports pricing of $1.25 per million input tokens and $4.25 per million output tokens.
Tools & repos
6 selected
Perfai Security
Perfai Security targets teams shipping vibe-coded apps by finding and fixing live vulnerabilities from a prompt. It is most useful as a quick security pass before exposing generated applications to real users.
iOfficeAI/OfficeCLI
OfficeCLI is an open-source single binary that lets agents read, edit, and automate Word, Excel, and PowerPoint files without installing Microsoft Office. It is a practical bridge for document-heavy enterprise workflows that need file-level automation rather than screenshots.
Autonomous Computer
Autonomous open-sourced complete build guides for personal and on-prem AI computers, including bills of materials, 3D-printable and CNC housing files, wiring diagrams, BIOS tuning, assembly photos, and setup notes for Ollama, vLLM, and llama.cpp. The configurations range from 2x RTX 5090 workstations to 4x RTX PRO 6000 Blackwell systems with 384 GB VRAM.
VoltAgent/awesome-design-md
awesome-design-md collects DESIGN.md analyses of popular brand design systems so coding agents can generate UI against a clearer style contract. It is useful for teams trying to reduce generic-looking agent-generated interfaces.
Timbal AI
Timbal AI packages agent building, workflow orchestration, and app creation in one stack. The clearest use case is teams that want to turn multi-step AI workflows into operational internal tools without stitching together separate agent and app layers.
Perplexity Computer GLM 5.2 orchestrator
Perplexity released a research preview of an orchestrator model for Perplexity Computer, adapted from GLM 5.2 and post-trained for its computer-use harness. The company claims near-frontier performance at 0.344x the cost of Opus.
Blogs worth your time
6 reads
Simon Willison breaks down the GPT-5.6 family
Willison gives the developer view of GPT-5.6: Luna/Terra/Sol pricing, 1M context, 128K output, benchmark caveats, and new API features such as programmatic tool calling, subagents, prompt cache breakpoints, and original-detail image inputs. Useful if you need to budget and benchmark the new models rather than just read launch claims.
Deep Agents Code on NVIDIA NemoClaw
LangChain shows a governed coding-agent blueprint using Deep Agents Code, NVIDIA NemoClaw, Nemotron 3 Ultra, deny-by-default networking, human approvals, and audit logs. The practical takeaway is how to run code agents near sensitive repositories without treating the model API as the only control surface.
NVIDIA’s iterative recipe for synthetic financial datasets
NVIDIA walks through a generation-deduplication loop that produced 502,536 unique financial news headlines across 13 categories using NeMo Data Designer, NeMo Curator, and Nemotron served via vLLM. The useful pattern is global semantic deduplication, farthest-from-centroid few-shot steering, and category reweighting for domain datasets that otherwise collapse into duplicates.
Aurora 1.5 extends Microsoft’s open Earth-system foundation model
Microsoft Research describes Aurora 1.5, adding 22 weather variables, hourly temporal resolution, and probabilistic ensemble forecasts to its open Earth-system foundation model. This matters for teams building weather, energy, transport, agriculture, and climate-risk products that need an extensible model rather than a closed forecast feed.
Apple studies self-reflective program search for long-context reasoning
Apple’s SRLM paper augments recursive long-context methods with uncertainty signals—self-consistency, trace length, and verbalized confidence—to choose better context-interaction programs. It is relevant for RAG and agent systems where simply stuffing more context does not reliably produce better reasoning.
Google SensorFM turns wearable streams into a health foundation model
Google Research introduces SensorFM, a wearable-health foundation model pretrained on more than one trillion minutes of sensor data from five million people. The post is useful for health-AI teams thinking about time-series foundation models, label-efficient adaptation, data infilling, and grounding personal health agents in longitudinal sensor data.
Funding & acquisitions
6 moves
Fundamentum Partnership
Fundamentum launched its third fund with a target corpus of ₹2,200 Cr, including a ₹400 Cr greenshoe option. The fund will target Series B startups in consumer tech, fintech, and AI with ₹100-150 Cr cheques, while Nandan Nilekani steps back from the GP role and remains anchor investor.
Ollama
Ollama raised a $65 Mn Series B led by Theory Ventures after growing to 8.9 Mn monthly developers and 176,000 GitHub stars. The company helps developers run open-weight AI models locally and also offers hosted access to larger models billed by GPU time.
Nurix AI
Mukesh Bansal-led Nurix AI plans to acquire Verloop.io for an undisclosed amount to expand from voice AI into unified voice and chat agents. Verloop.io brings chat automation capabilities, enterprise customers, and a platform automating more than 20 Mn customer interactions every month.
WizCommerce
Bengaluru-based WizCommerce raised ₹79.5 Cr, about $8.3 Mn, from existing investors including Blume Ventures, Alpha Wave, Peak XV, Z47, and 100X.VC. The company offers an AI-native sales and ecommerce SaaS platform for wholesale distributors and is estimated at a ₹443 Cr post-money valuation.
Hakimo
Hakimo raised a $12 Mn growth round led by existing investor Zigg Capital, taking total funding to $32 Mn. The Indian-origin physical AI company uses existing surveillance cameras for real-time security monitoring, forensic search, safety, compliance, and customer-experience workflows.
Gradium
Paris-based voice AI model startup Gradium reopened its seed round to new investors including Nvidia and has now raised $100 Mn in total. The company is building ultra-low-latency voice models and opening a Bay Area office to compete for AI talent.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.