The Software Efficiency Report | 2026 Week 44

Metric of the week

Environment Contention Ratio (ECR): Target Under 5%

Deep dive

Patching Faster Won’t Save You: Designing for a World Where Flaws Are Found at Machine Speed

Software Efficiency Metric of the Week

Environment Contention Ratio (ECR): Target Under 5%

What it is: The percentage of engineering time lost to waiting on shared test clusters, locked staging servers, or broken test data.

Why it matters: Shared staging environments quietly crush team throughput.

When three squads push to the same box, collisions are guaranteed:

• Squad A runs an experimental schema migration.
• Squad B gets random test failures and wastes two hours debugging them.
• Squad C pauses their ticket, context-switches to something else, and loses momentum.

Teams often try to manage this by setting up a booking calendar or putting someone in charge of environment slots. That does not fix the problem. It just puts an engineer in charge of supervising a traffic jam.

How to fix it:

1. Ephemeral preview environments: Spin up an isolated stack for each pull request, then destroy it automatically when the PR merges.
2. Database branching: Use copy-on-write snapshots so developers test against realistic data without overwriting each other.
3. Contract testing over giant end-to-end suites: Stop spinning up 15 microservices just to check an API response. Validate boundaries independently in CI.

The Formula: ECR (%) = (Total engineering hours blocked waiting on environments / Total delivery hours) * 100

Quick Example: A 6-person team logs 480 hours in a two-week sprint.

If developers spend 48 cumulative hours waiting for environment locks, debugging dirty data, or asking who broke the build:

ECR = (48 / 480) * 100 = 10%

You are burning 10% of your engineering budget just standing in line.

Track this metric for a single sprint. If your ECR is in double digits, you do not need more developers. You need to untangle your pre-production path.

Reader Poll

If your CI pipeline builds and tests in eight minutes, but your release sits in a manual QA queue for four days, do you actually have continuous delivery?

My take: Teams spend hundreds of thousands of dollars optimizing build runners, parallelizing test jobs, and bragging about shift-left practices. Yet when code is ready to ship, the entire delivery engine screeches to a halt so a human can manually click through a spreadsheet of test cases. Automated tests are supposed to give you the confidence to deploy. If you still require a multi-day manual regression freeze before every release, your test suite is not a safety net. It is just an expensive pre-flight ritual that nobody actually trusts.

How does your team decide code is safe enough to deploy to production?

• A) Full trust in automation: Unit, integration, and contract tests pass in CI, and code deploys directly to production or canary without manual sign-offs.
• B) Quick sanity check: Automated suites handle 95% of regression coverage, and the author spends ten minutes doing a quick smoke test before release.
• C) The manual regression freeze: Releases get batched into a multi-day testing freeze while QA manually clicks through a checklist on a shared staging server.
• D) Production is staging: We do not have automated tests or a dedicated QA team, so we ship it and wait to see if customer support gets flooded with tickets.

Does your automated test suite give you the confidence to ship on a Friday afternoon, or is your delivery velocity still held hostage by manual sign-offs?

Engineering Tip of the Week

Put an expiration date on temporary architectural shortcuts to stop technical debt drift

When a squad asks for permission to bypass a platform standard to hit an urgent product deadline, never grant an open-ended pass. Attach an explicit expiration date to the approval and schedule the remediation ticket in the very next sprint before greenlighting the shortcut. If you do not plan the cleanup work before shipping the hack, that emergency workaround quietly becomes permanent production architecture the moment the release goes live.

Technology Ecosystem Trends

Top Developments/Trends picks for this week Shaping Modern Engineering Efficiency and Operations

Data platforms chase the GPU clouds' business: Databricks' AI Runtime, managed serverless GPU training built straight into its data platform, reached general availability with a customer claiming a training run cut from seven days to 48 hours, a sign data platforms now want AI training spend that used to go straight to standalone GPU clouds. Source

Sovereign Infrastructure and Edge-Cloud Decoupling Enterprises are moving away from centralized, one-size-fits-all hyperscaler architectures toward sovereign multi-cloud deployments dictated by regional data regulations and local inference costs. For tech leaders, operational resilience now requires designing infrastructure layers that isolate core business logic from single-provider dependencies and enforce automated data-residency compliance. Source

Open Source and Linux: Addressing the "AI Slop" Pull-Request Surge Empirical industry findings published in late October 2026 show that while generative AI has cut development time, 25% of incoming pull requests to open-source repositories are now categorized as low-quality or non-functional synthetic code. Platform leaders must deploy mandatory contributor verification gates and automated test-coverage requirements before synthetic PRs exhaust core maintainer capacity. Source

Security and DevSecOps: Identity-Based Supply Chain and Cloud Secrets Hardening Widespread cloud credential compromises and poisoned distribution channels in October 2026 forced organizations to move beyond static scanning to runtime zero trust and identity-based secrets rotation. Technology heads are removing long-lived API tokens and developer credentials from CI pipelines, replacing them with short-lived, cryptographically attested workload identities. Source

AI/ML and Agentic AI: Agent Control Planes and Bounded Permissions As agentic systems move from pilot demos into multi-step production pipelines, engineering organizations are shifting focus from foundational model size to execution reliability and tool sandboxing. CTOs are implementing formal agent control planes that enforce least-privilege tool access, human-in-the-loop checkpoints for irreversible actions, and runtime evaluations inside CI/CD loops. Source

Embedded Systems and IoT: Mixed-Criticality Hypervisors and Long-Term Base OS Industrial edge deployments are adopting mixed-criticality Linux virtualization to run certified real-time control routines side-by-side with telemetry and AI inference engines on common hardware. For engineering heads in IoT and manufacturing, this standardizes updates via containerized over-the-air (OTA) pipelines without compromising underlying functional safety certifications. Source

An AI now decides what is safe for an agent to run: GitHub's Copilot Assisted Permissions reached general availability and uses an LLM to auto approve low risk agent actions while still stopping for anything riskier, a direct, shippable answer to research showing vague prompts push coding agents past their intended scope more often than anyone would guess. Source

PHP2Shell moves from warning to government outage: CISA confirmed active exploitation and gave federal agencies an October 29 patch deadline, the bug took down dozens of UK government services this week, and PHP's maintainers accused security vendor Chainguard of breaking the disclosure embargo by shipping patches six hours early, a messy capstone to a flaw that's already hit a meaningful share of the web. Source Source

AI becomes the default way to find software bugs: OpenAI and Wiz used frontier models to find the PHP2Shell flaw before it became a CISA emergency, Anthropic says Claude independently surfaced 493 unreported security issues across 25 open source projects including a bug that had hidden for 21 years, and Sam Altman told an Oracle conference crowd that AI is finding vulnerabilities "at a rapidly increasing rate," which is now as much a defender's tool as an attacker's head start. Source Source

Chinese labs keep setting the open weight pace: Moonshot AI shipped Kimi K3.1 Pro, a 2.8 trillion parameter open weight model it says cuts its own hallucination rate roughly in half on an internal benchmark, continuing a run of frontier scale open releases from Chinese labs that Western open weight efforts are still chasing on practical coding ability. Source

Tools, Resources and Communities | Worth Knowing

Open Source Tools

• Dagger: A programmable CI/CD engine that executes build pipelines inside containers using code instead of brittle YAML, guaranteeing that workflows run identically on a developer's laptop as they do on remote CI runners. Source
• RAUC (Robust Auto-Update Controller): An atomic update client for embedded Linux that manages dual-bank A/B system deployments with cryptographic verification, ensuring edge hardware recovers cleanly if an over-the-air field upgrade fails. Source
• Buildroot: A straightforward cross-compilation framework that generates minimal, reliable embedded Linux root filesystems, offering a lighter and faster alternative to complex build engines for dedicated hardware targets. Source

Commercial Platforms

• Bunnyshell: An Environments-as-a-Service platform that automatically spins up isolated, full-stack preview environments directly from pull requests and destroys them on merge, eliminating staging bottlenecks and queueing. Source
• Graphite: A developer workflow and review platform built around stacked pull requests, helping teams ship small, 200-line code reviews in rapid succession without waiting for earlier branches to merge. Source

Learning Resources – Interesting articles

• Embedded Artistry: A dedicated community and reference library for systems engineers covering hardware abstraction layers, embedded test-driven development, and production firmware architecture. Source
• Brendan Gregg's Systems Performance Hub: An essential reference for senior engineers on Linux kernel diagnostics, detailing profiling methodologies, eBPF tracing tools, and flame graph analysis for production latency issues. Source
• eBPF.io: The community hub and reference directory for eBPF technology, offering architectural documentation, tutorials, and real-world case studies for low-overhead kernel tracing and networking. Source

Deep Dive Article : Patching Faster Won’t Save You: Designing for a World Where Flaws Are Found at Machine Speed

Two things happened within hours of each other on October 21st. The PHP team released emergency patches for two critical PHP flaws (PHP2Shell), with public exploit analysis following almost immediately. That same day, a poisoned LiteLLM release (1.107.0) hit PyPI and Docker Hub, remaining live for about two hours before quarantine and it is long enough to reach production pipelines and harvest cloud credentials.

Neither gave anyone a planning window. And this isn’t a one-off: Google’s threat intelligence team found that half of AI-discovered vulnerabilities lead to remote code execution, and exploitation now often begins before a patch exists. I’m writing this because “scan, triage, patch in 30 days” no longer matches how attacks work. Here’s how to prevent, detect, and handle issues when they show up without warning.

What actually changed

Finding flaws got cheap. AI agents can read commits, diff patches, and work out exploits in a fraction of the time it used to take a person. Google’s Big Sleep agent found a SQLite flaw before it shipped in a release. The same capability is available to attackers, who can read a vendor’s patch and derive the exploit before most teams have finished the advisory.

The numbers show it. Google’s threat intelligence group tracked monthly vulnerability disclosures rising from about 5,000 in January to more than 10,000 by August this year. VulnCheck reports the median time from a CVE being published to it landing on CISA’s exploited list fell from 120 days in 2025 to 80 days in the first half of 2026. And for vulnerabilities that were exploited, Mandiant’s data put the average time-to-exploit at negative seven days in 2025, meaning attacks often began before a patch existed.

The tools you use to find flaws are targets. Security scanners and CI tooling run with elevated access: production credentials, private repos, secrets. That’s why attackers go after them. The March LiteLLM compromise was linked to credentials stolen in a compromise of the Trivy scanner, while the October incident exploited stolen GitHub tokens. Your scanner and release pipeline are part of your supply chain, so pin them, verify them, and limit what they can reach.

AI-written code reintroduces old bugs. Assistants produce what’s statistically likely, not what your policy requires. Veracode’s 2025 study of more than 100 models found that 45% of AI-generated code samples introduced security flaws. Attackers’ automated scanners love that.

Noise is rising too. Defenders are also drowning in low-quality, AI-generated vulnerability reports. Triage has become its own bottleneck.

Let us discuss now about how to prevent, how to get alerted and how to handle security flaws.

Prevent: assume the flaw exists and make it useless

You can’t patch your way out of a negative time-to-exploit. What you can do is limit what an exploited flaw gets the attacker.

Strip the environment. Say an attacker lands an RCE like PHP2Shell inside a container. On a traditional image they get a shell, curl, wget, cloud metadata endpoints, and a view of the internal network. On a distroless image running in a disposable sandbox (gVisor or a Firecracker microVM), there’s nothing to spawn, nothing to persist in, and nothing to scan with. The bug still fires, but the payload has nothing to work with. (A caveat for Python services: you can’t remove the interpreter, but you can remove everything around it.)

Lock down egress. Almost every exploit chain needs to phone home: web shell callbacks, credential exfiltration, command-and-control. Default-deny outbound traffic, allow only known endpoints, and enforce it below the application, at the firewall, network policy, or hypervisor layer. If the TCP handshake fails, the theft fails.

Add a release cooldown. LiteLLM 1.107.0 was caught in about two hours. A 48 to 72 hour hold on non-critical updates would have kept it out of most pipelines. Renovate has a minimumReleaseAge setting, Dependabot supports cooldowns, and tools like uv can exclude packages newer than a given date. Keep a fast path for genuine security fixes, so the cooldown doesn’t slow down an urgent patch.

Pin by hash. Version numbers alone aren’t enough. Use lockfiles with SHA-256 hashes (uv.lock, poetry.lock, pip with require-hashes), pin container images by digest, and pin CI actions by commit SHA. After the March LiteLLM incident, teams with lockfiles pinned to a known-good version were unaffected.

Limit agent and scanner permissions. If an agent only summarizes tickets, its credentials shouldn’t allow deleting data, writing files, or sending traffic out. Treat anything an agent reads (emails, webhooks, web pages) as untrusted input, and put a deterministic rules engine between the model’s tool call and the actual execution. If a call breaks the schema, it doesn’t run.

Keep static analysis in the loop, and cut the noise. SAST (Semgrep, CodeQL, SonarQube) and SCA (Black Duck, Snyk, Dependabot, OSV-Scanner) still catch real problems, especially in AI-generated code. Add reachability analysis where your tools support it, so you fix the vulnerable code that actually runs. Be aware that SCA tools matching against known-vulnerability databases won’t flag a brand-new malicious release, since nobody has catalogued it yet. Pair them with behavioral package scanning, such as Socket.

Alert: know before your vendor tells you

Monthly scans guarantee a blind spot. Layer a few real-time signals:

• CISA KEV: Pipe the catalog into your security channel. A KEV listing means exploitation is confirmed in the wild, so it’s a prompt to act, not to schedule.
• EPSS: It estimates how likely a flaw is to be exploited. A rising score is a reason to look closer and reprioritize, but it isn’t proof that exploit code exists.
• Behavioral package monitoring: Socket, OSV-Scanner, and similar tools can flag a dependency that suddenly requests network access, runs obfuscated base64, or adds a post-install hook.
• Runtime detection: Falco or Tetragon (eBPF-based) can alert the moment a container spawns an unexpected shell or opens an outbound connection it shouldn’t.
• Canary tokens: Seed fake, monitored credentials in vector stores, document stores, and config files. If something external tries to use one, you know data has been read and exfiltrated, and which system leaked it.
• Researcher feeds: Independent researchers and threat teams often publish technical analysis quickly, sometimes before advisories are complete. Follow Project Zero, watchTowr, Qualys, Datadog Security Labs, and the communities you trust on Mastodon and Bluesky.

Handle: the part most teams skip

Prevention and alerting get all the attention, but in practice the first hours after an alert decide how bad it gets. Build for speed here too.

Answer “are we affected?” in minutes. Generate an SBOM on every build (Syft, CycloneDX) and store it somewhere queryable. When the next advisory drops, you should be able to ask “do we run litellm 1.107.0 anywhere?” and get an answer immediately, without scanning live systems. Re-check stored SBOMs against new advisories continuously, so you learn about a flaw in something you shipped last quarter.

Triage by exposure, not just severity. Combine four questions: is there a working exploit, is it being used by attackers, does it touch an asset we expose, and is a fix available? An internet-facing PHP server with public exploit code goes to the top of the list. A 9.8 in an unreachable internal library does not.

Have containment moves ready. When you can’t patch in time, buy time: a WAF or virtual patch rule, a feature flag, network isolation of the affected service, or temporarily taking an endpoint offline. Decide in advance who’s authorized to do each one.

Write the supply-chain runbook before you need it. If a malicious package may have run in your CI or production: freeze deployments, identify every build that pulled the affected version during the exposure window, and rotate every credential those environments could reach. Short-lived, scoped credentials make that rotation fast. Long-lived keys scattered across pipelines make it a week-long project.

Make patching a pipeline, not a project. Automated tests, canary rollouts, and one-click rollback let you ship an emergency fix in hours. If every patch needs a manual change window, you’ll lose the race.

Rehearse. Run a tabletop exercise: “a malicious version of our model gateway was published an hour ago.” You’ll find the gaps (who has access to rotate that key, who owns that service) before an attacker does.

A checklist you can use (valid this week, as it gets updates)

☐ Dependencies pinned by hash; containers pinned by digest; CI actions pinned by SHA
☐ Release cooldown configured (48 to 72 hours) with a fast path for security fixes
☐ Scanners and CI tooling pinned, least-privileged, and treated as supply chain
☐ Production images are distroless or minimal; shells and network tools removed
☐ Code-executing agents run in ephemeral sandboxes, destroyed after each task
☐ Default-deny egress with an explicit allowlist
☐ SAST, SCA, and a behavioral package scanner running in CI
☐ SBOMs generated per build, stored, and queryable
☐ CISA KEV feed and runtime detection wired to an on-call channel
☐ Canary tokens in context stores and config
☐ Incident runbook written, credential rotation tested, tabletop done this quarter

The bottom line

When discovery is automated, human review can’t be the only checkpoint. The teams that do well won’t be the ones who patch fastest in every case. They’ll be the ones whose environments make an exploited flaw boring: nothing to run, nowhere to connect, and a clear plan for the hours after the alert.

Analyze this about what you’re seeing in your own pipelines: Which of these is your biggest gap right now: prevention, detection, or response?

Sources: Source Source Source Source Source Source

Technology Ecosystem Weekly News Digest – Top Picks

Cloud and Platform News

AWS this week:

• AWS Continuum, its tool for finding, validating and proposing fixes for code vulnerabilities, reached general availability in the two US regions where it launched, pairing a vulnerability's business impact with a reproduced exploit before opening a pull request for a human to approve. Source
• Amazon SageMaker AI now offers training session APIs that let teams write custom training loops, including reinforcement learning reward signals, while AWS still handles GPU allocation and fault recovery, billed per token processed. Source
• Amazon Bedrock added structured, schema compliant JSON outputs across every Claude model, cutting the custom parsing and retry logic teams usually build for document extraction and support ticket routing. Source
• Amazon ElastiCache Serverless for Valkey can now migrate a self-managed Valkey or Redis OSS instance with zero downtime and no third party replication tools, and the migration nodes are free for the first 180 days. Source
• AWS launched Agent Registry, a managed catalog where publishers register MCP servers and agent tools with searchable metadata so orchestrators can discover them at runtime instead of hard coding every tool at build time. Source

Azure this week

• Application Gateway's WAF policies can now be scoped at the route level instead of only per domain, so a strict policy can sit on a login or checkout path underneath a looser, profile wide baseline. Source
• Microsoft Foundry's cost management now breaks out Claude spend by model and by input versus output tokens instead of lumping it into generic SaaS cost, closing a chargeback gap for teams running Claude through Azure. Source
• Azure quadrupled the default limit on standard storage accounts per region per subscription, from 250 to 1,000, and made the increase self-service instead of a support ticket. Source
• Network Watcher's VNet flow logs can now be filtered by IP, port, protocol and direction before they're written, cutting log volume and cost for security and compliance review. Source
• Azure Backup added item level recovery for VMs on Premium SSD v2 and Ultra Disk, so a single file or folder can be restored without pulling back the whole disk. Source

Google Cloud this week.

• Google Cloud launched a self-managed VMware Cloud Foundation offering on dedicated bare metal across 22 regions, giving customers full root access to ESXi, vCenter and NSX Manager instead of the fully managed VMware Engine, and it cites unaudited vendor figures of up to 40 percent lower server costs. Source
• BigQuery picked up a bundle of general availability features at once: native Git support inside BigQuery Studio, a Data Engineering Agent tied to Google Drive, conversational analytics in Connected Sheets, and a new AI.COUNT_TOKENS function. Source
• Google is shutting down the legacy Chronicle BigQuery export dataset for customers below its Enterprise Plus tier, forcing a migration to customer-managed exports or a tier upgrade. Source
• AlloyDB for PostgreSQL reached general availability on IAM group based authentication and a four-level ScaNN index tree for vector search. Source

Nvidia and Oracle are betting that agentic AI needs its own networking. The two expanded their Vera Rubin deployment on Oracle Cloud Infrastructure to what they're calling multi-gigawatt scale, brought the Vera CPU to OCI, and pitched Oracle's Acceleron networking as hardware isolation for running untrusted AI-generated code. Oracle says OpenAI already trains and serves GPT-6 Astra on an OCI supercluster with more than 100,000 Blackwell GPUs. This is the companies' own joint announcement, so treat the capacity claims as theirs rather than independently measured. Source

Open-Source and Linux Ecosystem News

CNCF and Kubernetes this week.

• A follow-up to the eMudhra certificate authority compromise covered here two weeks ago: patches cascaded out this week across cloud native infrastructure once the downstream fix was ready, with Cilium, cert-manager, Envoy, Prometheus, Crossplane, Dapr, SPIFFE/SPIRE, OpenTofu, Atlantis, Dex, Kagent, K8up, Trivy and Kube-OVN all shipping releases for CVE-2026-111943, which traces back to a Go toolchain bump needed so crypto/x509 rejects chains through the bad intermediates. Source Source Source
• The Kubernetes project stood up a Workload Conformance Working Group to write a shared standard for how applications should behave at runtime, since existing conformance testing only checks the cluster, not the software running on it. Source

Tesco is looking to walk away from VMware. The UK grocery chain posted a job listing for an edge platform engineering lead whose remit includes migrating store level virtualization from VMware to Proxmox, a move that lines up with its public lawsuit accusing Broadcom of pricing a 2026 renewal roughly 175 percent above what its original 2021 contract allowed. Tesco says a full exit realistically can't happen before the end of 2027. Source

Fedora Linux 45 missed its release date for the second time. The distribution was due October 20, then October 27, and is now targeting November 3 after testers found a bug where encrypted Stratis storage pools can fail to initialize at boot. Three more issues, including broken SSH passphrase handling after an upgrade, are still working through blocker review. Source Source

The creator of CentOS and Rocky Linux built a working kernel from scratch in eleven days, with an AI agent writing nearly all of it. Gregory Kurtzer's "Ginux" isn't a Linux fork. It has its own memory manager, scheduler and filesystem code but implements the Linux ABI closely enough to boot Rocky Linux 10 and run unmodified binaries under QEMU. He used a single persistent OpenAI Codex session to produce roughly 139,000 lines of retained code across more than 1,000 commits, and he's released it as a research prototype with no GPU support and no security audit. Source

DevOps, Platform Engineering and SRE News

GitHub this week. At GitHub Universe, the company took Copilot Fusion, a layer that routes a coding task to whichever model or combination of models fits it best, to general availability, and says its top effort tier scores within two points of Claude Opus 5.5 Max on a coding agent benchmark at roughly half the cost. Alongside it came a new agent first Copilot app, parallel multi-repo agent sessions in preview, a higher code review tier, and a change to Enterprise billing that lets customers pay for Copilot through AI credits instead of a separate subscription. Source Source

JFrog this week.

• AppTrust, JFrog's governance layer on top of Artifactory, added a pre-mapped compliance framework for the EU's DORA rules covering change management, component integrity, patching, resilience testing and incident response, plus "AI Verification Policies" that have an LLM check whether a release actually matches its ticket before it ships. Source
• The same week, JFrog extended its shadow AI detection to developer laptops, so it can flag unapproved AI skills, plugins or MCP servers installed locally, not just in the pipeline. Source

OpenTofu gave its planner a way to reason about values it used to just give up on. Version 1.13 adds convert and assumenotnull functions so module authors can hint at a value's type or guarantee it won't be null, letting more count and for_each expressions resolve during tofu plan instead of stalling. Terraform shipped the same idea first, and OpenTofu matched its syntax rather than inventing its own. Source

Pulumi wants provisioning a GPU cluster to look like provisioning anything else. Its new provider wraps Nvidia's open source AI Cluster Runtime, a library of validated GPU infrastructure recipes tested across EKS, GKE and AKS on hardware from A100s to GB200s, into a single resource that gets the normal Pulumi preview, diff, drift detection and destroy lifecycle instead of being provisioned by hand. Source

Security and DevSecOps News

Stolen GitHub tokens are now being laundered through Actions' own logs. Wiz found a tool, nicknamed Gheist, that takes a stolen token, spins up a throwaway branch, commits a workflow that encodes every repository secret and prints it to the run log in a way that dodges GitHub's usual masking, grabs the log over the API, and deletes the run within seconds. It's hit more than 850 accounts and over 13,500 planted branches so far, including private repositories and self-hosted runners, where it can reach host files directly through Docker. Source

A Japanese cloud hosting breach turned into real financial fraud, not just a data leak. GMO Internet Group found implants in its FX trading desktop cloud service, and attackers used stolen trading credentials to actually place trades in some customers' accounts, confirmed independently by broker OANDA Japan. GMO refused a roughly $50 million extortion demand, one customer's data has already been leaked in retaliation, and the company is now offering a $1 million bounty, paid only on an arrest. Source

A Microsoft 365 token theft campaign hit 79 organizations by replaying stolen sessions in bulk. Huntress tracked attackers stealing post-login session tokens through fake sign-in pages and device code phishing, then reusing them en masse from a narrow range of IP addresses, racking up 497 successful logins against more than 200,000 failed attempts. In nineteen of the affected organizations, the attackers swapped out existing MFA methods within three hours of getting in. Source

South Korean police made what they're calling the country's first arrest tied to AI-driven hacking at scale. A former penetration tester allegedly used an autonomous AI attack framework called ARTEX AI to breach more than 90 organizations, including hospitals and schools, stealing roughly 8.2 million records and sending out 38 separate ransom demands. Six other suspects remain under investigation. Source

AI/ML and Agentic AI News

An OpenAI API spec becomes an open standard: OpenAI donated its "Open Responses" LLM API format to the Linux Foundation's Agentic AI Foundation, and it already has real backers in AWS Bedrock, Databricks, Nvidia, Red Hat and Hugging Face, though Anthropic and Google have conspicuously not adopted it yet. Source

The two biggest hyperscalers' earnings both point the same way: AI spending keeps climbing, and it's starting to show up as real cloud revenue, not just promises. Microsoft's fiscal first quarter put Azure revenue up 48 percent year over year to $33.2 billion, with capital spending roughly doubling to $39.4 billion. Meta's third quarter revenue rose 26 percent to $64.6 billion, and it narrowed its full year infrastructure spending guidance upward to $135 to $145 billion while unveiling Muse Wonder, which it calls its first frontier model. Source Source

Nvidia wants to be the dashboard for AI data centers that have outgrown human readable alerts. Its new Minerva platform ingests telemetry across GPUs, networking and storage, builds a live map of what's actually running, and uses machine learning to collapse alert floods into a single root caused incident, with a natural language assistant on top. Early adopter IREN is already using it, and Dell, Lenovo and Supermicro plan to bundle it. Source

Replit built an agent whose job is to be the on call engineer. Its new Monitoring Agent investigates a production alert, finds a likely cause, drafts and validates a fix, and, if a customer has opted into auto publish, deploys it straight to production before emailing a summary of what changed. Replit flags that fixes involving database migrations are riskier to leave on autopilot, the kind of caveat that will matter to anyone actually turning this on. Source

Anthropic handed its agent packaging format over to a neutral foundation. It donated Agent Skills, the folder based format it created for giving agents reusable procedural knowledge, to the Linux Foundation's Agentic AI Foundation, the same body that already holds the Model Context Protocol. Anthropic says more than 70 platforms have adopted it already, including ChatGPT, Cursor and GitHub Copilot, with close to 280 million installs. Source

Embedded Systems and IoT News

In a follow-up to PHP2Shell, the critical PHP bugs covered here the past two weeks, industrial automation vendors are now checking whether a web server flaw became a factory floor problem. After PHP2Shell was added to CISA's known exploited list, Rockwell Automation and Schneider Electric both published advisories auditing their own product lines for exposure, a reminder that PLC and HMI web interfaces often embed the same general purpose stacks that run ordinary websites. Rockwell says it hasn't yet confirmed any affected products; Schneider is still pointing customers to a separate bulletin for specific product lines. Source Source

The FCC blocked two devices for carrying the wrong chip, not for being the wrong brand. It barred a wireless video transmitter and a dashcam because both use HiSilicon chips, a Huawei subsidiary on its covered list, the first enforcement under a rule that took effect this month extending the device ban to anything that merely contains a covered company's silicon, even under an unrelated brand. This is reported by a single outlet so far. Source

A kernel filesystem proposal is trying to fix data corruption in places you can't just add a second hard drive. A new RFC called BeamFS, aimed at single drive systems exposed to radiation like satellites and radiotherapy equipment, stores error correcting codes in every block so corrupted reads get silently repaired or cleanly fail instead of quietly returning wrong data. It's an early proposal, not a merge candidate, and its own documentation warns of data loss and crashes. Source

Closing Note

This week's issue highlights a growing challenge for engineering teams: keeping software delivery predictable while security risks, infrastructure changes and development tools continue to evolve.

From shared test environment bottlenecks and temporary architectural shortcuts to the latest developments across cloud platforms, Kubernetes, DevOps, AI agents and embedded systems, the common thread is operational discipline. The deep dive explores how to prepare for vulnerabilities discovered at machine speed, with practical steps to strengthen CI/CD pipelines, limit exposure and improve incident response.

If you want to understand where your engineering delivery system needs improvement, start with TuskerGauge. It is a free assessment covering CI/CD, testing, infrastructure, security, observability, SRE, deployment safety and engineering practices.

Find the gaps in your engineering delivery system

Already know where the bottlenecks are? Tusker90Pro helps turn those findings into a practical 90-day improvement roadmap focused on delivery flow, reliability, release management and engineering operations.

Build your 90-day improvement roadmap

The goal is to identify the constraints, reduce operational risks and build a software delivery system that teams can trust.

If you are dealing with a specific engineering delivery challenge, write to contact@stonetusker.com. We can discuss the problem, the constraints and practical steps to improve it.