Today’s lead · TechCrunch
Anthropic prototypes cryptographic proofs for model identity
Anthropic’s FlashProve points at a practical trust layer for hosted models: prove a response came from the claimed model without exposing weights. It is still research, not a deployment plan, and early commentary flags real overhead questions builders should watch before betting on it.
Top signals
5 morePruna open-sources P-H3-Image from MiniMax-H3
Pruna’s P-H3-Image is a useful example of video-model reuse turning into a lighter open image model. The headline is not just quality; it is operational fit: 20 GB instead of 143 GB, 8-step generation, Diffusers support, and native ComfyUI support.
LangSmith BYOC expands from AWS to Azure and Google Cloud
LangSmith’s BYOC move is aimed at the compliance gap between cloud SaaS and self-hosting. AWS is now GA, while Azure and Google Cloud are beta, letting traces, evals, datasets, prompts, secrets, and agent deployments stay in the customer cloud account.
Google warns LLMjacking can burn six figures in a day
LLMjacking turns leaked or stolen AI credentials into a fast billing disaster, not just an abuse nuisance. Google and Mandiant observed more than $100,000 in fraudulent charges in a single day, with some organizations reaching $1 million before access was revoked.
AI data centers are using converted jet engines for power
The AI power crunch is getting weird: some data center builders are using retired aircraft engines rebuilt to burn natural gas. The appeal is speed, not elegance. The trade-off is worse efficiency and pollution than conventional gas turbines, making this look like a bridge strategy.
Apple pushes dating and romantic AI companion apps to 18+
Apple’s new App Store policy tightens the path for minors into dating apps and AI companions with sexual or romantic content. For builders, the bigger signal is classification pressure: companion apps now have to treat age ratings and content descriptors as product architecture.
Tools & repos
5 selectedJustVugg/colibri
Colibri is a C inference engine for frontier MoE models that tiers experts across GPU, RAM, and NVMe, targeting CUDA and Vulkan on existing hardware.
AgentKit by Scalekit
AgentKit is an MCP gateway for production agents, wrapping 30,000+ tools across 600+ apps with auth, scopes, logs, custom connectors, and event triggers.
max-sixty/worktrunk
Worktrunk is a Rust CLI for managing Git worktrees, aimed at parallel AI agent workflows where multiple branches need clean isolation.
MachineMail
MachineMail gives agents real email inboxes over MCP or API, with custom domains, inbound trust reports, and a free tier of 3,000 messages per month.
Testifly
Testifly pitches autonomous AI QA for small teams: connect a web app, then it discovers flows, writes tests, self-heals, and explains failures with video.
Blogs worth your time
5 reads
Browser agents are mostly a tooling problem
LangChain’s browser-agent writeup is useful because it compares seven toolsets, then shows how traces and evals drove concrete fixes: fewer model turns, less context, and faster waits.
StateBench tests whether models can actually track changing state
Ai2’s StateBench separates easy “latest value” tracking from harder chains of updates. The practical warning: frontier models still degrade on long, stateful tasks as dependencies get deeper.
Simon Willison stress-tests weight-space hype on a laptop model
Willison reproduces part of the weight-space programming paper with Qwen 3.8 27B locally, finds it impressive, and usefully reins in the “source code is over” narrative.
IBM’s Palette shows how small models can ship structured creative output
Palette is a good pattern for enterprise generation: split planning from code, train on synthetic reasoning, then put a checker and repair loop around the model.
DeepMind frames Antigravity as a cheminformatics workbench
DeepMind’s Antigravity demo is strongest when it keeps the agent as orchestrator and lets deterministic chemistry tools handle calculations. The caveat is explicit: designs remain hypotheses.
Community discussions
2 threadsn8n users debate what survives when agents can write the glue code
The thread’s useful split is deterministic workflows versus agentic one-offs. Commenters argue n8n still earns its place when jobs must run predictably, with agents handling the messy parts inside.
Grok Bot usage audits expose the hidden tax of rereading context
The post is a concrete reminder that agent cost often hides in orchestration defaults. Timer-based routines and long chats can spend most tokens rereading history instead of doing work.
Funding & acquisitions
3 moves
Fal raises $1.3B as inference demand surges
Fal’s round shows how hot inference infrastructure remains, especially around media models. The caution is margin pressure: TechCrunch notes inference providers face high compute costs even as demand and valuations climb.
CrowdStrike buys Fiddler AI and Dreadnode for agent security
CrowdStrike is buying AI observability and testing pieces as enterprises deploy agents faster than security policy can catch up. Fiddler becomes Falcon Observe; Dreadnode feeds testing results into Falcon Guardian.
PowerUp Money raises $10M for AI-led wealth management
PowerUp is pushing beyond mutual fund advice into AI-led planning and managed wealth. The opportunity is clear in India’s SIP growth; the hard part is keeping digital advice suitable, transparent, and compliant.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.