The New Stack

The New Stack has launched a new ebook, sponsored by HPE, with 10 practical checks for Kubernetes platforms handling AI workloads and agents.

Discover KubeVirt's future: confidential computing, cutting-edge AI hardware, and modular, pluggable virtualization stacks.

Vercel has given Marketplace vendors until January 2027 to offer safer authentication and says OIDC support for deployments is rolling out.

DeepInfer lets developers self-host V4.1 Flash with DeepSeek’s own optimizations, but stops short of a complete production stack.

Mandated internal platforms were twice as likely to deliver their goals. Yet optional adoption now leads, and AI agents are changing the question.

OpenAI’s cheaper GPT-6.1 Terra posts big coding gains, but its own safety tests show the AI agent bypassing restrictions in nearly 30% of runs.

MiniMax says M3.1 Flash cuts agentic task costs by 70%. In three hands-on tests against M3, it came out cheaper, but slower and no more accurate.

Anthropic’s S-1 reveals models under six months old drove over 80% of Q3 revenue, sharpening a debate about developers’ pull to the newest models.

Ex-GitHub CEO Thomas Dohmke’s startup Entire launches a GitHub alternative that swaps pull requests for Trails, an eval-driven agentic code review system.

Ramp built its own AI agent harness, Laminar, instead of buying one. LangChain’s Harrison Chase sees a bigger shift in how agents get built.

Two critical PHP flaws, found with help from frontier AI, put WordPress, Drupal, and Joomla at risk. OpenAI wants a new playbook for the flood to come.

The S-1 puts Claude Code at the center of how Anthropic grows inside companies, and says Claude wrote more than 90% of contributions to Anthropic's codebase in September.

AWS made Continuum generally available after it scored 89% on the CyberGym-E2E code security benchmark. A closer look at what that number really shows.

Started at Owner.com last month, tsrs is now in Cloudflare’s hands, and the tricks it uncovered are making Microsoft’s own compiler faster too.

Things aren't always what they seem at KubeCon. Especially at the Museum of Illusions. The New Stack and Dynatrace are hosting a reception for the cloud native community, just a short walk from the Salt Palace Convention Center. Get your name on the list: www.eventbrite.com/e/night-at-t...

OpenAI’s cheaper GPT-6.1 Terra posts big coding gains, but its own safety tests show the AI agent bypassing restrictions in nearly 30% of runs.

What happens to your MFT environment when your most experienced admin walks out the door? Our upcoming ebook tackles the knowledge problem nobody writes down. Pre-register and get it first: bit.ly/47TIhTa

MiniMax says M3.1 Flash cuts agentic task costs by 70%. In three hands-on tests against M3, it came out cheaper, but slower and no more accurate.

Anthropic’s S-1 reveals models under six months old drove over 80% of Q3 revenue, sharpening a debate about developers’ pull to the newest models.

Ex-GitHub CEO Thomas Dohmke’s startup Entire launches a GitHub alternative that swaps pull requests for Trails, an eval-driven agentic code review system.

Mandated internal platforms were twice as likely to deliver their goals. Yet optional adoption now leads, and AI agents are changing the question.

Static sandboxes cannot contain dynamic AI agents. Discover why effective governance demands runtime observability and rapid containment.

The S-1 puts Claude Code at the center of how Anthropic grows inside companies, and says Claude wrote more than 90% of contributions to Anthropic's codebase in September.

Stop reading AI code line by line. Reviewing architectural decisions instead of diffs fixes the PR bottleneck while keeping teams aligned.

Machine intelligence may soon read your telemetry first. Who checks its work?

How do you investigate an AI-driven incident when the identity provider has no record of how an agent gained access? Tomorrow, Shawn Petty, CCO at CloudBolt, and TNS Host Viktor Farcic will examine how automation changes the identity and audit trail problem. Don't miss it: bit.ly/4686mF3

OpenTelemetry and interoperability help teams standardize telemetry, govern AI agent behavior, and measure cost per business outcome.

Agents should hand engineers a fix, the company says, and teams should still check the work.

Kubernetes won the orchestration wars, but dev and ops teams still pull in opposite directions. HPE's David Estes thinks AI may ease the strain.

OpenAI reports dramatic improvements in legacy software modernization, but it's keeping the tests behind its biggest claim private.

Two critical PHP flaws sent some sites offline. The fight over Chainguard’s early patch release may outlast the bugs themselves.

OpenAI is donating Open Responses, its spec for a common LLM API, to the Agentic AI Foundation. Here’s who supports it and who’s still missing.

If anyone with a Claude subscription can build it, it isn’t worth much. The value has moved to understanding what users actually need.

ElevenLabs’ Eleven Realtime bundles voice agent speech, reasoning and turn-taking into one service. Here’s what developers trade away for that simplicity.

Both companies want to make their AI chips more accessible to PyTorch developers. It’s still hard, though.

Agent Knowledge Foundation bundles retrieval, ontology, and prediction models for enterprise agents. Nvidia is counting on Snowflake, Databricks, and Palantir to decide who gets to see what.

Google’s Agent Compute Engine, which already powers Gemini, gives every AI agent its own hardware-isolated microVM. A preview arrives later this year.

PHP2Shell’s chaotic disclosure sparked a row. Behind the drama, maintainers warn that AI is piling new security work onto already stretched teams.

Letta’s new FrontierMemory benchmark tests whether AI agents learn from experience, and the harness matters as much as the model behind it.

Two critical PHP flaws, found with help from frontier AI, put WordPress, Drupal, and Joomla at risk. OpenAI wants a new playbook for the flood to come.

Partners will pre-build DSX POD containers with seven Vera Rubin racks and liquid cooling, cutting data center setup from months to as few as 14 days.

Haiku 5.5 missed just four of 1,600 fields. Two came from a date it shouldn’t have trusted

Vercel is pushing Marketplace vendors off long-lived API keys toward OIDC by 2027 — but its own GitHub Actions deploys still rely on a static token.

Ramp built its own AI agent harness, Laminar, instead of buying one. LangChain’s Harrison Chase sees a bigger shift in how agents get built.

Microsoft’s speech model starts transcribing before callers finish talking. But Phonely says the hard part of voice AI is the silence in the middle.

Swif.ai is expanding Linux device management so IT can say yes to developers. The pitch: AI teams get Linux, and security gets proof of control.

Is Kubernetes ready for AI agents? Discover how sandboxing, security guardrails, and agent substrates address unpredictable AI workloads.

Anthropic’s next coding model may infer developer intent. Research shows coding agents often go beyond scope, and that’s much harder to fix than it sounds.

Swif.ai is expanding Linux device management so IT can say yes to developers. The pitch: AI teams get Linux, and security gets proof of control.

Kubernetes scale isn't just nodes. Explore how control planes hit churn, state, and node limits, and how architectural seams fix them.