The developer stories enterprise IT leaders actually read.
Volume 343, Issue 3 · 14 articles
A Dark Reading roundtable says CISOs should fix non-human identity visibility in the SDLC first, then use AI to cut testing and patch costs.
A Dark Reading roundtable found AI is complicating software development security but also offers a way forward.
Cisohive founder Renee Guttmann says security teams are now sought-after collaborators and should push scanning deeper into development, including for user developers outside IT.
BlueFlag Security CISO Bobby Bauer calls identity the most open attack vector, citing 144 non-human identities per human and a Cloud Security Alliance finding that 68% of organizations cannot tell agent actions from human ones.
Panelists recommend short-lived credentials and full agent activity logging, and say AI is driving the cost of finding and fixing flaws toward zero.
AI is now finding real vulnerabilities at high volume, leaving engineering teams overwhelmed by valid fixes rather than fake reports.
Google paused vulnerability submissions to its open source bug bounty program after a surge of automated reports, its second crackdown this year.
LeadDev reports a shift from bogus AI reports to a high volume of legitimate findings that each need triage and a fix. curl creator Daniel Stenberg says reports arrive at four to five times the 2024 rate, forcing a month-long pause.
He advises teams to run AI scanners on their own code, while managers should avoid overloading teams with findings.
A practical guide to Unix timestamps: units, conversions, time zones, and avoiding precision, leap-second and Year 2038 bugs.
This beginner-friendly guide explains that a Unix timestamp counts seconds from the epoch, 1970-01-01 00:00:00 UTC, by convention, and that it does not guarantee a clock is correct or that two events get different values.
It covers the units a timestamp may use, from seconds to nanoseconds, and why the surrounding API must define which.
It walks through converting timestamps, handling time zones and storing them in databases, and explains pitfalls including precision loss, leap seconds and the Year 2038 overflow on 32-bit systems, with examples in common languages and tools.
Schools still teach on 1991's Turbo C++ under DOSBox; GCC with Code::Blocks is a free, standards-compliant replacement.
Many schools and colleges still teach C and C++ using Turbo C++ 3.0 from 1991, run through the DOSBox emulator because 64-bit Windows cannot run 16-bit DOS programs.
The author explains the difference between the language standard, the compiler and the IDE.
He shows how Turbo's pre-standard dialect deprives students of the STL, std::vector, std::string and modern practice, and teaches vendor extensions as if they were the language.
He recommends GCC with Code::Blocks, which works the same on Windows and Linux, as a free and legally clear replacement.
Agentic AI shifts distributed systems from orchestrating services to orchestrating runtime decisions, which brings new failure modes.
Part of InfoQ's Agentic AI Architecture series, this article compares microservices, which split up functionality behind fixed APIs, with agentic systems that split up decisions made at runtime.
Agents bring non-deterministic execution paths driven by context, reasoning and tool selection, making decision quality a reliability concern alongside latency and availability.
The author warns that production failures often show up as a quiet decline in reasoning quality rather than explicit errors.
He recommends bounded autonomy, evaluation loops and deterministic fallback paths.
Google's TorchTPU and AWS's TorchNeuron bring native PyTorch to TPUs and Trainium, though top performance still takes work.
At PyTorch Conference North America, Google open-sourced TorchTPU, a native PyTorch stack for TPUs built with Meta that removes JAX runtime dependencies, while AWS released TorchNeuron in public beta for Trainium2 and Trainium3.
Both aim to reduce dependence on NVIDIA's CUDA ecosystem.
Basic migration can need only small code changes, such as switching to the neuron device, and both support eager mode and torch.compile.
The New Stack cautions that neither is a silver bullet, and reaching top-tier performance on non-NVIDIA hardware still takes tuning.
How to scale Spring Batch across a small cluster using the existing relational database for coordination instead of Kafka or RabbitMQ.
This InfoQ article shows how to run fault-tolerant, distributed Spring Batch jobs using only the relational database Spring Batch already needs, rather than remote partitioning over a message broker.
The pattern sizes work to the live node count, assigns a master for each execution, and claims partitions with compare-and-set writes so none is lost, while heartbeats and two-phase recovery handle node failure.
Transferable work must be idempotent, and non-transferable work fails rather than moving.
The author says a broker is still the better choice for very large clusters or sub-second latency.
An ActiveState survey finds AI coding gains are offset by rework costing at least 139 hours per engineer per year.
LeadDev reports on an ActiveState survey of over 300 security, platform and engineering leaders at companies with 1,000+ employees.
43% say at least 10% of engineering time goes to fixing or revisiting AI-assisted code, a floor of 139 hours, or 17 working days, per engineer a year, which for a 100-person team equals about seven engineers' annual capacity.
The report cites DORA and Copilot research showing similar downstream bottlenecks.
It also finds only 31% have formally quantified the cost of unmanaged open source risk.
Why Linux clocks drift and jump, how NTP keeps them honest, and why elapsed time needs a monotonic clock rather than wall-clock time.
Your computer's clock can drift, jump backward and disagree with other servers, and this guide explains why.
Linux offers CLOCK_REALTIME for wall-clock time and CLOCK_MONOTONIC for measuring elapsed time, since the wall clock can be stepped during corrections; the lesson is to use wall time for when something happened and monotonic time for how long it took.
It explains how NTP synchronizes clocks despite network delays and faulty sources, how to inspect status with timedatectl, and why scheduling something for 9 AM tomorrow is harder than waiting 86,400 seconds because of time zones and daylight saving.
At PyTorch Conference, TorchTPU and TorchNeuron extend PyTorch to non-NVIDIA chips as the project strains under its own growth.
Steven Vaughan-Nichols reports from PyTorch Conference, where Google's TorchTPU and Amazon's TorchNeuron backends let existing PyTorch workloads run on TPUs and Trainium.
Modular's Chris Lattner said the modeling layer is unified but kernels, compilers and low-level software remain fragmented, and described Modular as an LLVM-like stack for AI chips.
Maintainers said more than 600,000 GitHub projects build on PyTorch and about 150 pull requests arrive daily, straining testing and releases.
Edward Yang said training teams care most about time to a usable checkpoint, which calls for better profiling and smaller, composable building blocks.
A Harvard study of 700+ firms finds AI agents lift code volume about 30% but resolved work only 12%, as review becomes the bottleneck.
InfoQ reports on a Harvard study that used Jellyfish data covering roughly 300 million work events at more than 700 firms.
AI coding assistants produced small gains in lines, commits and pull requests, while agents raised them 20-30%, but resolved Jira issues rose only 12% with agents and not at all with assistants.
The researchers estimate engineers spend just 30-40% of their time writing code.
More AI-generated code shifts work to reviews, testing and deployment, which limits overall gains unless those stages also improve.
Video explainer: Jira manages tasks, bugs and sprints, while Confluence is the place for documentation, notes and requirements.
This TechnologyAdvice video, produced and hosted by Caleb Kinchlow, compares two Atlassian tools that project teams often use together.
Jira is built for managing project work, including tasks, bugs, sprints and issue tracking, while Confluence provides a central place for documentation, meeting notes and project requirements.
The video helps teams decide which tool fits their needs, or how to use both.
TechnologyAdvice notes that some vendors may pay it for web traffic or sales opportunities.
Stack Overflow's own 2026 survey shows developers visiting less because of AI, as the company pivots to enterprise, hiring and data licensing.
The Stack Overflow 2026 Developer Survey of over 30,000 developers found 64% visit the site less for simple questions because of AI, and 27% contribute less.
LeadDev traces the company's response since its post-ChatGPT 'code red'.
That includes AI Assist, Stack Internal (formerly Stack Overflow for Teams), licensing its data for AI training starting with Google, a jobs partnership with Indeed, a revived Developer Story profile, and a Stack Overflow for Agents beta.
The article asks what happens to the developer knowledge commons as answers move into chatbots.
Kaseya's CTO says AI agents rewrote a storage layer in six weeks, and clear specifications and domain knowledge now matter most.
Kaseya CTO Pratik Wadher says his team used Anthropic coding agents to help write a spec, reviewed it with critique agents, then generated and tested code.
With this approach they re-architected the storage back end to work with anything from ZFS to S3 in six weeks.
He argues domain expertise, critical thinking and clear writing will matter more than syntax, and that documentation becomes critical.
He says the spec, not the code, is what teams will check into GitHub, though he has seen variation in generated output.
More from Developer · All IT section newsletters · This week