Today’s lead · thehackernews.com
A hostile webpage can break out through an agent host
Microsoft researchers detailed AutoJack, an exploit chain against AutoGen Studio where a browsing agent that opens an attacker page can reach a privileged local MCP WebSocket and trigger host command execution. The vulnerable code shipped in the 0.4.3.dev1 and 0.4.3.dev2 PyPI pre-releases, while the fix is currently in GitHub main at commit b047730 rather than a patched PyPI release. This caught my attention because it is the exact production failure mode agent builders keep underestimating: untrusted web content, localhost assumptions, MCP-style tool surfaces, and code execution all living too close together.
Top signals
19 moreCodex is becoming a teachable agent runtime
OpenAI’s Codex now has Record & Replay, which captures a user workflow once and turns it into an inspectable, editable skill the agent can reuse later. In the same wave of builder chatter, Codex is being shown handing threads between local and remote hosts, driving Google Cloud settings through its in-app browser, and sustaining large subagent-heavy sessions, which makes it feel less like a coding assistant and more like an operating environment for software work.
Copilot CLI now talks back and second-guesses
GitHub says new Copilot CLI features are generally available: developers can speak to Copilot using on-device speech-to-text models, and a built-in Rubber Duck agent uses a second model to surface blind spots. The important bit for me is not the voice input itself, but the pattern: coding agents are starting to ship with native critique loops instead of leaving review discipline entirely to the human.
Anthropic’s model ban became a frontier-lab stress test
TechCrunch reports that the White House forced Anthropic to restrict Fable 5 and Mythos 5 after national security concerns involving Mythos access and alleged Fable safeguard bypasses, leaving the models unavailable while the company scrambled to comply. The episode matters because export controls are now colliding directly with frontier model release operations, customer access, cyber-defense use cases, and the trust layer labs need to run global developer ecosystems.
ASML’s EUV mystery hits the AI chip chokepoint
U.S. officials are reportedly pressing ASML over concerns that EUV-related tools or components may have reached China, while ASML says no EUV machine has ever been shipped there. This is worth tracking because EUV lithography is one of the deepest bottlenecks behind Nvidia-class AI chips, and even partial leakage of know-how could shorten China’s path toward leading-edge semiconductor capacity.
Google published the TPU scaling receipts
Jeff Dean pointed to a new IEEE Micro paper covering Google’s training supercomputers from TPU v2 through Ironwood, with claims including roughly 30x better TFLOPS per watt, 10x HBM capacity and bandwidth per node, and a jump from 256-chip TPU v2 pods to 9,216-chip Ironwood pods. For infra people, this is a useful rare look at how an accelerator architecture survived the transformer era while scaling power, resilience, interconnects, and pod size.
Baseten’s valuation jump shows inference is still scorching
Baseten is reportedly close to raising $1.5 billion at a headline valuation of $13 billion, only months after a $300 million Series E valued it at $5 billion. The detail I care about is that the inference layer is still being funded as a core control point for LLM apps: fast serving, cost control, model routing, and open-source fallback selection are becoming infrastructure moats.
Elastic buys into AI-native debugging
Elastic has reportedly agreed to acquire Deductive AI for up to $85 million, bringing in a startup focused on using AI to catch and resolve software bugs. This fits the bigger devtools shift: as AI-written code volume rises, observability vendors need agents that can move from alert to root cause to remediation without making humans manually stitch the loop together.
OpenAI adds research heft and policy cover
OpenAI is bringing in Noam Shazeer, a co-author of the original Transformer paper and former Gemini lead, alongside Dean Ball, who will lead a new Strategic Futures policy team focused on catastrophic risk, recursive self-improvement, labor impact, and lab-government relations. The pairing is a clean read on OpenAI’s IPO-era priorities: keep frontier technical credibility high while building a much tighter policy interface than rivals facing government pressure.
Chinese open models keep winning router traffic
GLM 5.2 is getting notable developer attention, including an Amp plugin with zero data retention and U.S./Western inference, while a separate OpenRouter chart circulating today claims Chinese models have moved from about 1% to more than 50% of token consumption in under a year. I would not overfit to one chart or one model review, but the production signal is real: cheap, fast, good-enough models are reshaping routing decisions faster than brand-name frontier labs would like.
Reliance wants agents inside India’s phone network
Reliance announced Jio Call Agent for its 500-million-user telecom network, plus AI tasking inside MyJio and an ambient home display called TeleFrame. The engineering angle is distribution: instead of another standalone assistant app, Jio is trying to make transcription, summarization, bookings, plan changes, and household reminders native to calls, apps, and connected devices.
Claude Code workflows are turning into agent swarms
The Claude Code discourse today clustered around parallelism: one post says Anthropic engineers are running 100-plus agents with self-improving loops, another describes a Google Cloud developer splitting a build across API, pipeline, and dashboard agents, and a third shows Claude working through an Unreal Engine MCP server. The through-line is that serious users are moving from one chat window to planned, inspectable, multi-agent execution across tools.
Depth Anything 3 gets the GGUF treatment
Depth Anything 3 now reportedly runs as pure C++/ggml with a self-contained GGUF, avoiding Python, PyTorch, and the CUDA toolkit at inference time. That matters because the ggml/GGUF ecosystem is moving beyond text LLMs into useful vision primitives, making local multimodal pipelines easier to ship on constrained hardware.
Cerebras tees up Gemma 4 multimodal access
Cerebras is offering early access to Gemma 4 on its hardware and pairing it with a 24-hour hackathon backed by Google DeepMind and Cerebras. The claim to watch is speed: if Cerebras can make a new multimodal Gemma feel instant enough for builders, it becomes another serious option in the model-routing toolbox.
Trio ships a developer layer for live physical operations
IoTeX and MachineFi Lab introduced Trio, a world model for physical operations, with Trio-Retina available for developers via pip install and Trio-Lumen for operators who want to ask camera feeds questions in natural language. The interesting part is the product shape: turning video from warehouses, factories, and retail floors into structured scene state that agents can monitor, predict on, and eventually act against.
VLC’s builder is open-sourcing remote-control infrastructure
Jean-Baptiste Kempf, best known for VLC, has raised $5 million for Kyber, an SDK for synchronizing video, audio, sensor data, and control inputs with minimal latency across remote devices. I like this because physical AI will need boring, robust plumbing for robots, drones, and fleets long before the high-level agents look magical.
Snap moves costly AI video work outside the house
Snap is spinning out its internal generative AI video team into Dotmo, a separate company focused on models for interactive gaming experiences, with CTO Bobby Murphy personally investing and Snap retaining equity and a technology license relationship. This is a useful market signal: even large consumer platforms are looking for ways to keep exposure to AI video upside without carrying all of the compute burn on the core P&L.
Shadow AI has become an identity problem
The Hacker News piece frames the next enterprise AI security issue less as employees pasting secrets into chatbots and more as unsanctioned agents connected to Salesforce, Snowflake, GitHub, Slack, MCP servers, browser extensions, and service accounts. That is the right mental model for production teams: agents are non-human actors with credentials, permissions, owners, lifecycles, and blast radius.
YC’s batch says agents need better sandboxes
TechCrunch’s YC Demo Day roundup shows the Spring 2026 cohort leaning heavily into defense tech, robotics, AI infrastructure, developer tools, and agents, with some valuations reportedly pushing past $175 million. The startup I’d actually want to inspect is Arga Labs, which is building digital twin environments for testing AI agents before their code reaches production.
AI networking is not just copper versus optics
SemiAnalysis argues that scaling GPU clusters should lift demand for both copper and optical interconnects, because they solve different reach, power, cost, reliability, and bandwidth constraints. The practical takeaway is simple: as model clusters grow, data movement becomes a first-order systems problem, not a single-component bet.
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 movesClaude Code Artifacts
Preview and share your coding work live as it happens
Zernio WhatsApp API
One API for WhatsApp: messaging, calling, and AI agents
Midjourney Scanner
60 second ultrasound-based full-body scanner that beats MRI
Firecrawl Research Index
An index for agents pushing the frontier of AI/ML research
API to MCP
Turn any API into an MCP server for AI agents
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.