Today’s lead · thehackernews.com
A single webpage can turn AutoGen Studio into an RCE path
Microsoft researchers disclosed AutoJack, an exploit chain where an AI browsing agent loads a hostile page, the page reaches AutoGen Studio’s local MCP WebSocket service, and an unauthenticated command path can spawn a host process. The vulnerable surface was not in the stable 0.4.2.2 PyPI release, but did ship in 0.4.3.dev1 and 0.4.3.dev2; the hardening is in GitHub main at commit b047730, not yet a PyPI build. This is the production-agent failure mode I care about most: localhost trust, browser context, MCP, and tool execution collapsing into one boundary.
Top signals
21 moreCodex learns reusable skills from recorded workflows
OpenAI added Record & Replay for Codex, letting a user demonstrate a recurring workflow once and convert it into an inspectable, editable skill. That moves coding agents closer to persistent team automation rather than one-off chat sessions, which is exactly where developer-agent UX needs to go.
Claude Code turns agent sessions into shareable live artifacts
Anthropic rolled out Artifacts in Claude Code for Team and Enterprise plans, so a coding session can publish an interactive page such as a PR walkthrough or living project dashboard. I like this because it treats the agent’s intermediate work as a collaborative object, not just terminal output that disappears into scrollback.
Claude’s robodog test shows agent autonomy moving into hardware workflows
Anthropic reran Project Fetch with Claude Opus 4.7 operating through Claude Code and reports it completed the software-side robodog tasks roughly 20x faster than last year’s best human team using Opus 4.1. The model still failed at precise physical ball retrieval, but the pattern is notable: tasks that first needed humans plus AI are becoming mostly model-driven.
OpenAI is training for alignment that survives new domains
OpenAI published research on reinforcement learning for broadly beneficial behavior, using realistic scenarios across health, education, science, law, engineering, economics, and other domains to improve traits like honesty, corrigibility, fairness, and epistemic humility. The interesting claim is generalization: small amounts of beneficial-trait data mixed into post-training improved many independent safety and alignment evals and made models harder to steer toward harmful behavior.
Agentic QA is becoming the missing loop for AI-written code
TestSprite 3.0 is positioning itself as a verifier that drives the running product: parallel exploration agents, per-agent video replay, backend testing, UI drift auto-heal, dynamic variables, and machine-readable failure bundles for coding agents. The sharp part is the framing: if agents can write code overnight, the verifier has to live in the same terminal/IDE loop and feed fixes back to the agent.
Natural-language rewards are entering agent RL pipelines
OpenPipe’s ART stack is getting attention for RULER-style rewards, where criteria are written in plain English and an LLM grades each agent trajectory for GRPO training instead of relying on hand-coded scoring functions. The linked docs also point to W&B serverless RL, instant checkpoint deployment, and examples like email search, 2048, LangGraph, MCP, and game environments.
GLM-5.2 has open-weight engineers asking for better serving paths
The GLM-5.2 chatter is unusually technical: Sebastian Raschka highlights MLA, DeepSeek Sparse Attention, and a new IndexShare trick that reuses sparse-attention token indices across layers to make 1M-token inference cheaper. Practitioners are already asking which harnesses and providers can run it well, which is the real test for any strong open-weight model.
Cerebras puts multimodal Gemma 4 on a 1,500-token-per-second path
Cerebras says Google DeepMind’s Gemma 4 is now running as its first multimodal model on the platform at 1,500 tokens per second, claiming a 15x speedup over the nearest comparable model. For agent loops that need see-reason-act-retry cycles, this kind of latency change matters more than another small benchmark delta.
Liquid AI ships tiny multilingual retrieval models with millisecond latency claims
Liquid AI introduced LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M for multilingual and cross-lingual retrieval across 11 languages, claiming end-to-end retrieval latency as low as 1.5ms with its enterprise stack. Fast retrieval models are still one of the highest-leverage parts of RAG systems, especially when the model layer keeps getting cheaper and more routable.
An open 9B extraction model is closing in on frontier document parsing
Vik Paruchuri’s team is open sourcing a 9B model for structured document extraction that scores 90.2% on their benchmark, close to Gemini 3.5 Flash at 91.3% and ahead of NuExtract3 at 81.5%. The practical details matter here: 9.5s p50 timing and JSON-schema compliance make it directly relevant for document AI pipelines.
Qualcomm’s NPU compiler looks much stranger under the hood
A reverse-engineering writeup on Qualcomm’s NPU compiler claims it silently downgrades model weight precision, uses the HiGHS linear-programming solver for memory placement, and can produce a 33x DDR-traffic difference across chips with identical reported VTCM. Edge AI deployment is mostly compiler behavior now, and this is the kind of undocumented machinery that explains why real-world NPU performance is so hard to predict.
Architect Labs wants AI to design and verify chips end-to-end
Architect Labs launched with a $24M seed round led by Kindred Ventures and a founding team that says it has taped out 80+ production chips and worked across Meta AI silicon, Intel neuromorphic chips, Anthropic, xAI, and Google DeepMind. The company says some AI-generated chip designs are headed to leading-edge tapeout later this year, which makes this more interesting than the usual chip-startup pitch deck.
Google’s TPU paper puts hard numbers on eight years of accelerator scaling
Jeff Dean pointed to a Google paper covering TPU v2 through Ironwood, including a 10x increase in HBM capacity and bandwidth per node, 100x peak node performance, 3,600x supercomputer performance, and roughly 30x better TFLOPS per watt. The pod-scale jump from 256 chips in TPUv2 to 9,216 in Ironwood is a useful reminder that training progress is as much systems engineering as model architecture.
Amazon is openly considering Trainium racks as a chip business
AWS is in early talks to sell Trainium AI chips for third-party data centers, with Andy Jassy previously framing the internal chip effort as a potential $50B standalone run-rate business. If Amazon actually sells racks outside AWS, it becomes a more direct Nvidia competitor instead of just a cloud provider with custom accelerators.
Baseten’s next round shows how hot the inference layer still is
Baseten is reportedly close to raising $1.5B at a headline valuation around $13B, only months after a $300M round at $5B. The deal structure may be split-priced, but the signal is clear: model routing, serving cost control, and open-model inference are still drawing enormous capital.
Elastic is buying into AI SRE with DeductiveAI
Elastic has reportedly agreed to acquire DeductiveAI for up to $85M, bringing in a startup focused on using AI to catch and resolve software bugs and system failures. This fits the broader post-codegen bottleneck: observability vendors want agentic remediation, not just dashboards that tell humans what broke.
Frontier labs are recruiting for both models and governance
OpenAI is adding Transformer co-author Noam Shazeer and policy operator Dean Ball ahead of its expected IPO path, while AlphaFold lead John Jumper says he is leaving Google DeepMind for Anthropic after a recharge break. The talent moves say a lot about where frontier labs think leverage lives now: core model intuition, scientific credibility, and direct policy/governance capacity.
Anthropic’s model-access fight is turning into a governance test case
TechCrunch’s coverage says the U.S. government forced Anthropic to pull Fable 5 and Mythos 5 over national-security concerns after reported guardrail bypasses, while security researchers have pushed back that similar jailbreak issues exist elsewhere. The important bit for builders is not the branding drama; it is that frontier-model availability may now change abruptly because of policy, nationality, and access-control constraints.
Reliance is trying to make AI native to the phone network
Reliance announced Jio Call Agent, an AI assistant that can join calls to transcribe, summarize, and perform tasks like booking cabs or reservations, with a launch planned for Jio’s 500M-plus users. It also showed AI features in MyJio, a TeleFrame home display, sector-specific services, and broader infrastructure plans, making this one of the more serious distribution plays for ambient AI.
OpenAI’s o3 Deep Research gets a rare-disease case study
OpenAI and Boston Children’s Hospital used o3 Deep Research in work discussed around a new NEJM AI paper on helping diagnose rare genetic diseases in children. The reason I’d still track this as an engineer is that medical diagnosis is a brutally high-stakes retrieval, reasoning, and uncertainty-calibration workflow, not just another chatbot demo.
Databricks’ Omnigent gets a search partner for grounded agents
Nimble says it is the launch web-search partner for Databricks’ Omnigent, an open-source AI agent meta-harness, exposing search as a configurable provider for orchestrated agents. It is a small integration story, but it captures the larger move from picking one model to composing model, search, tools, cost, and accuracy policies into an agent stack.
Tools & repos
3 selectedDeusData/codebase-memory-mcp
High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies.
google-research/timesfm
TimesFM (Time Series Foundation Model) is a pretrained time-series foundation model developed by Google Research for time-series forecasting.
palmier-io/palmier-pro
macOS video editor built for AI
Blogs worth your time
0 readsNo long-form read was strong enough to recommend today.
Funding & acquisitions
5 movesUpstream
The inbox designed for humans and agents
Honestly
See what Reddit and TikTok honestly think about your product
Jesse
Stop building Apollo/Clay lists. Search the live internet.
Tabstack Dev Tools
Ditch your scraper. Make one API call with any tool.
Elvin
Proactive AI that finds and finishes work before you ask
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.