📬 Linux 7.0's PostgreSQL Crisis, OpenClaw's Triple CVE, TigerFS

⚡ TL;DR

• Linux 7.0-rc7 released April 5 with AI security documentation and WiFi fixes, but an unresolved PostgreSQL performance crisis (throughput halved on Graviton4) clouds the mid-April stable release
• OpenClaw hit by three critical CVEs in 2026, including CVE-2026-25253 (CVSS 8.8 RCE) affecting 40,000+ exposed instances, plus credential leakage and privilege escalation flaws
• TigerFS mounts PostgreSQL as a filesystem via FUSE/NFS, letting AI agents navigate databases with ls and cat using ACID transactions
• OpenHands reaches approximately 72% SWE-bench Verified, ships modular Software Agent SDK refactored into four independent packages
• KDE Plasma 6.7 scheduled for June 16 with global push-to-talk, print queue app, and Wayland blur protocol support

🏆 Top Story

Linux 7.0-rc7 Released with AI Security Docs, WiFi Fixes, and an Unresolved PostgreSQL Performance Crisis

• Linux 7.0-rc7 shipped April 5, 2026, with stable Linux 7.0 targeted for April 12 (or April 19 if an rc8 is needed). The release includes new security documentation designed to help AI agents submit better bug reports, and fixes the Ath11k/Ath12k WiFi driver performance regression that existed since 2019. About half of the patches address drivers (GPU, networking, USB, sound). The most consequential development around rc7 is not in rc7 itself: an AWS engineer reported on April 3-4 that PostgreSQL throughput on Linux 7.0 drops to 0.51x on AWS Graviton4 instances. The root cause is Linux 7.0's removal of PREEMPT_NONE as the default preemption model for modern CPU architectures (ARM64, x86, PowerPC, RISC-V, s390, LoongArch). PostgreSQL's spinlocks assumed non-preemptive scheduling, in which lock holders would not be interrupted mid-operation. Under the new PREEMPT_LAZY model, threads holding locks get preempted, causing other threads to waste CPU cycles spinning. A patch was proposed to restore PREEMPT_NONE, but the kernel maintainer, Peter Zijlstra, countered that PostgreSQL should adopt Restartable Sequences (RSEQ) instead. No clean resolution exists before stable release. Linux 7.0 will power Ubuntu 26.04 LTS (April 23) and Fedora 44 (April 14). This regression is a distribution-level concern.
• Source: Phoronix (PostgreSQL) | 9to5Linux

PostgreSQL throughput just dropped in half on AWS Graviton4. And the kernel developer who could fix it says it's not his problem.

Welcome back. Today we're talking about Linux 7.0. A kernel release is about to ship with a performance regression that no one would like to fix. And a philosophical fight about who should even try.

Plus, OpenClaw got hit with three critical CVEs in three months. A Princeton professor mounted PostgreSQL as a filesystem. And KDE Plasma 6.7 is bringing push-to-talk to Linux desktops.

Today's theme: people making decisions, and other people paying for them.

Linux 7.0-rc7 shipped on April 5th. Stable release is targeted for April 12th. Maybe April 19th if Linus adds an rc8.

On paper, this looks routine. Driver patches for GPU, networking, USB, sound. New security documentation to help AI tools submit better kernel bug reports. A WiFi driver fix for a regression hanging around since 2019.

Half the patches are drivers. Normal cycle stuff. Quiet week.

But then an AWS engineer posted benchmark results on April 3rd and 4th.

Elephant in the room: PostgreSQL throughput on Linux 7.0 drops to 0.51x on Graviton4 instances.

That's not a small dip. That's half. Half your database performance is gone because the kernel changed how it handles thread scheduling.

Think of it like renovating a building's foundation while tenants are still inside. The architect says the new structure is stronger. The tenants say their doors don't close anymore. Both are telling the truth.

Now, Graviton4 is ARM64. And ARM64 instances on AWS are increasingly the default for cost-optimized deployments. This isn't a niche edge case but mainstream production infrastructure.

I go deeper on these shifts every week, what they mean for your architecture, and the decisions underneath.

So what actually happened?

Linux 7.0 removed something called PREEMPT_NONE as the default preemption model for modern CPU architectures. ARM64, x86, PowerPC, RISC-V, s390, LoongArch. All of them.

If you're not deep into kernel internals, here's what matters.

Preemption controls when the operating system can interrupt a running thread. Under the old model, PREEMPT_NONE, if your thread held a lock, the kernel mostly left it alone. The thread would finish, release the lock, and other threads could proceed.

The new model, PREEMPT_LAZY, is more aggressive. It can interrupt a thread even while it's holding a lock.

PostgreSQL uses spinlocks. A thread tries to grab a lock. If it can't, it spins in a tight loop, burning CPU cycles, waiting for the holder to finish.

Under the old model, that spin was short. The holder finished quickly because nothing interrupted it.

Under the new model, the holder gets preempted mid-operation. Other threads keep spinning. CPU cycles burn for nothing. Throughput collapses.

A patch was proposed to restore PREEMPT_NONE. Kernel maintainer Peter Zijlstra pushed back. His position: PostgreSQL should adopt Restartable Sequences, called RSEQ, a newer kernel mechanism designed for exactly this kind of coordination.

Translation to human language: imagine a bathroom with one key hanging on a hook by the door.

Under the old rules (PREEMPT_NONE), if you grabbed the key and went in, nobody would drag you out mid-business. You'd finish, hang the key back, and the next person would go. The line moved fast because everyone inside got to finish.

Under the new rules (PREEMPT_LAZY), a manager can tap you on the shoulder mid-handwash and say, "Your break's over, come back later." You leave with the key still in your pocket. Now everyone in line is standing there, checking the hook every two seconds, burning time. Nobody can go in. You're somewhere else, not even aware you're still holding the key.

That's a spinlock under preemption. PostgreSQL's threads are the people in line. They don't go do something else while waiting. They stand at the hook and check. Check. Check. Check. Burning CPU cycles instead of doing actual work.

The old system was simple: don't interrupt the person inside. The new system is smarter in theory (the manager can reallocate people more efficiently), but it assumes everyone in line will go do something useful while waiting. PostgreSQL doesn't. It just stands there and spins.

Zijlstra's fix (RSEQ) is basically teaching PostgreSQL to leave a phone number on the hook rather than stand in line. "Call me when it's free." Good idea. But PostgreSQL has been standing in line for 30 years. You can't rewrite that habit in two weeks before the building opens.

Both sides are technically right. Zijlstra is right that fewer preemption models simplify kernel maintenance. Cleaner code, fewer edge cases. PostgreSQL is right that its spinlock design worked for decades under well-understood assumptions.

The problem is timing. PostgreSQL can't ship an RSEQ implementation before Linux 7.0 goes stable. That's weeks away.

And here's the distribution angle. Ubuntu 26.04 LTS ships on April 23rd. Fedora 44 ships April 14th. Both may ship with Linux 7.0. Ubuntu LTS means five years of support. Five years with this regression baked in, unless someone carries a downstream fix.

Now, while the kernel community argues about preemption models, the application layer is dealing with its own growing pains.

OpenClaw, the most-starred repository on GitHub with over 346,000 stars, just collected its third critical CVE in 2026.

CVE-2026-25253. CVSS 8.8. Remote code execution through a CSRF flaw that auto-connects via WebSocket and leaks the auth token. Over 40,000 instances were internet-exposed at the time of disclosure. 63% assessed as remotely exploitable.

Then CVE-2026-33575. Credential leakage through pairing setup codes. If someone grabs your pairing code from a chat log, a screenshot, or a log file, they get the gateway credentials. Full access.

Then CVE-2026-33579. Privilege escalation through the /pair approve command. CVSS 8.6. An attacker with basic operator.pairing scope can silently approve admin-level device pairing requests.

Three distinct vulnerability classes. Three months. A project that went from zero to mass adoption in sixty days.

The security review process couldn't keep pace with adoption. Sound familiar? Same pattern as the kernel regression. Growth outrunning the governance structures meant to manage it.

On the other end of the spectrum, something interesting happened this week.

Michael Freedman, co-founder of TigerData and professor at Princeton, released TigerFS. It mounts PostgreSQL databases as directory trees using FUSE on Linux and NFS on macOS.

Every file is a database row. Writes are ACID transactions. Multiple agents and humans can read and write simultaneously with full transactional guarantees.

His thesis, and I'm quoting him directly: "Agents don't need fancy APIs or SDKs. They love the file system."

It's counterintuitive. We've spent years building specialized APIs for everything. Freedman's bet is that the filesystem, the interface agents already know, is the right abstraction. Worth watching.

And quickly: OpenHands hit roughly 72 percent on SWE-bench Verified, refactored their SDK from a monolith into four independent packages. KDE Plasma 6.7 arrives June 16th with global push-to-talk and a print queue app. Elive, a Debian-based distro, came back after more than six years with 32-bit support and an anti-e-waste mission.

Step back for a second.

The PostgreSQL regression is technically clean. The kernel team simplified the preemption model. That's good engineering. PostgreSQL's spinlocks were built on assumptions that held for decades. That's also good engineering.

So who pays for the transition?

The missing column here is the transition cost.

The kernel team measures architectural cleanliness. PostgreSQL measures throughput. Nobody is measuring the cost that lands on distribution maintainers. They have to choose: carry a downstream patch or ship a known regression. And nobody is measuring the cost to production operators who discover this after upgrading.

I go much deeper on this every week, what it means for your architecture and the decisions ahead.

And it's the same pattern with OpenClaw. The project outpaced its security review process. The cost of that gap doesn't land on the developers who wrote the code. It lands on the 40,000 operators running exposed instances.

Transition cost. Growth cost. The bill always goes to someone who wasn't in the room when the decision was made.

So what should you do?

If you run PostgreSQL on ARM64, test on Linux 7.0 before you upgrade. Don't wait for your distribution to decide for you.

If you're a distribution maintainer, decide now. Carry the PREEMPT_NONE patch or pass the regression through. Your answer says something about who you think should bear the transition cost.

If you run OpenClaw, audit against all three CVEs. Not just the latest one. All three.

And keep an eye on TigerFS. Whether filesystems become the default agent interface is still open. But the thesis is strong enough to track.

🧩 Open Source News

OpenClaw Hit by Three Critical CVEs in 2026, Including One-Click RCE Affecting 40,000+ Exposed Instances

• OpenClaw's security posture continues to deteriorate despite its status as the most-starred GitHub repository (346,000+ stars as of April 2026). Three CVEs have been disclosed in 2026. CVE-2026-25253 (CVSS 8.8, RCE): a CSRF flaw auto-connects via WebSocket, leaking the auth token. Over 40,000 instances were internet-exposed at disclosure, with 63% assessed as remotely exploitable. Fixed in 2026.1.29. CVE-2026-33575 (Information Disclosure): long-lived gateway credentials embedded in pairing setup codes allow attackers who obtain leaked codes (from chat history, log files, or screenshots) to recover and reuse shared gateway credentials. Fixed in 2026.3.12. CVE-2026-33579 (Privilege Escalation, CVSS 8.6 v4.0 / 8.1 v3.1): the /pair approve command fails to forward security scopes. An attacker with basic operator.pairing scope can silently approve device pairing requests for operator.admin scope. Fixed in 2026.3.28. The pattern is clear: a project that reached mass adoption in 60 days is now cycling through critical security disclosures at a pace that challenges users' ability to stay patched.
• Source: NVD | SentinelOne | DEV Community

TigerFS Lets AI Agents Navigate PostgreSQL with ls and cat

• TigerFS, built by Michael Freedman (co-founder and CTO of TigerData, formerly Timescale), mounts PostgreSQL databases as directory trees using FUSE on Linux and NFS on macOS. Every file corresponds to a database row. Writes are ACID transactions. Multiple agents and humans can read and write concurrently with full transactional guarantees. The project ships with Claude Code skills for agent interaction. Freedman's thesis: "Agents don't need fancy APIs or SDKs, they love the file system." Released under MIT license. The approach inverts the typical database-as-API pattern, instead presenting structured data through the filesystem interface that AI agents already know how to use.
• Source: InfoQ | GitHub

OpenHands Reaches ~72% SWE-bench Verified, Ships Modular Software Agent SDK

• OpenHands achieves approximately 72% resolution rate on SWE-bench Verified, with top configurations using Claude Sonnet 4.5 with extended thinking. On January 28, 2026, the project published the OpenHands Index, benchmarking 9 language models across standardized evaluation criteria. The SDK has been refactored from the monolithic V0 architecture into four independent Python packages: openhands.sdk (core agent abstractions, event system, MCP protocol), openhands.tools (tool implementations), openhands.workspace (execution environments supporting local and Docker), and openhands.server (deployment). The redesign moves from mandatory Docker to optional sandboxing with LocalWorkspace as the default for lower friction. V0 is deprecated as of April 2026. The modular design enables production-grade agent deployments with stateless, event-sourced components and typed, swappable interfaces.
• Source: OpenHands Blog | GitHub

OpenScreen Hits 17,000+ GitHub Stars as Free Screen Studio Alternative

• OpenScreen, created by Siddharth Vaddem, gained 2,573 stars in a single 24-hour period and has reached approximately 21,500 stars total. The MIT-licensed tool offers per-window and full-screen recording, automatic and manual zoom with adjustable depth, microphone plus system audio capture, custom backgrounds, annotations, webcam overlay, and no watermarks or subscriptions. Version 1.3.0 (released April 2) adds undo-redo, localization (Chinese, English, Spanish), and webcam picture-in-picture mode. OpenScreen positions itself as a free alternative to Screen Studio (which charges $29/month or $108/year, with no lifetime option currently available). It targets developers who create product demos and tutorials.
• Source: GitHub

TermHub Gives AI Agents Programmatic Control Over iTerm2 and Windows Terminal

• TermHub (GitHub: duo121/termhub) provides a native terminal control gateway for AI agents. On macOS it controls iTerm2; on Windows it controls Windows Terminal. Capabilities include managing tabs, panes, and sessions, sending commands, pressing keys, and capturing output programmatically. A built-in session checkpoint loop ensures that when an AI agent sends a command, it captures only the new output produced after that command, not historical terminal output. Checkpoints are session-scoped, so two AI agents can operate different sessions in parallel without conflict. The project also ships an SDK (createTermhubClient) for building custom programmatic clients.
• Source: GitHub

Aethyr Edge Node Brings ML-KEM-768 Post-Quantum Cryptography to ESP32-S3

• Published April 5 by CNX Software. Aethyr Research released open-source firmware for the ESP32-S3 implementing formally verified ML-KEM-768 (NIST FIPS 203) for post-quantum key encapsulation, BLAKE3 integrity hashing, and XChaCha20-Poly1305 encryption. The firmware boots in 2.1 seconds with post-quantum handshakes completing in 35ms on ESP32-S3-WROOM-1 modules with 8MB PSRAM. NIST mandates quantum-resistant security for federal systems by 2035, and governments across the US, EU, Canada, and Australia require post-quantum upgrades to begin by 2026. Running ML-KEM-768 on a $5 microcontroller demonstrates that post-quantum cryptography no longer requires high-end hardware.
• Source: CNX Software

RISC-V ISA Manual Gets April 5 Snapshot: Zalrsc and Zaamo Extensions Formally Listed

• The RISC-V ISA manual published a snapshot release tagged riscv-isa-release-be23e0c-2026-04-05. The Zalrsc (Load-Reserved/Store-Conditional) and Zaamo (Atomic Memory Operations) extensions are now formally listed. These are documentation maintenance snapshots, not ratified specification events. The Zaamo extension enables microcontroller-class implementations to use atomic primitives without requiring cache-dependent LR/SC instructions. Part of a multi-release sprint on the ISA manual repository.
• Source: [GitHub](https://github.com/riscv/riscv-isa-manual/releases

🐧 Linux News

KDE Plasma 6.7 Coming June 16, 2026 with Push-to-Talk, Print Queue App, and Standardized Wayland Blur Protocol

• KDE Plasma 6.7 ships June 16, 2026, with Beta 1 on May 14 and Beta 2 on May 28. Major features: a global push-to-talk function (all microphones muted until a configured key is pressed), a light/dark mode panel toggle for instant switching, a full-featured print queue viewer application, custom sound theme installation from downloaded files, and support for the ext-background-effect-v1 Wayland protocol for standardized blur effects. Additional changes include excluding windows from screen recording, revamped VPN settings, improved multi-GPU detection, and rounded Breeze selection highlights in Dolphin, Okular, and KMail. The push-to-talk feature addresses a common complaint from gamers and video conferencing users about applications defaulting to open microphones.
• Source: KDE Blogs

GNOME 51 "A Coruna" Scheduled for September 16, 2026

• GNOME 51 is named after A Coruna, Spain, the host city for GUADEC 2026 (July 16-21). Release schedule: Alpha on June 27, Beta on August 1, Release Candidate on August 29, and final release on September 16, 2026. GNOME 50 is the current stable release, shipping in Ubuntu 26.04 LTS and Fedora 44. No feature list has been published yet for GNOME 51. The schedule follows the established six-month cadence.
• Source: Tux Machines | GNOME Release Calendar

Elive 3.8.50 LTS Returns After Seven Years with Enlightenment, 32-Bit Support, and OpenRC

• Elive "Retrowave" 3.8.50 LTS released March 30, 2026, marking the first stable release in over six and a half years. Built on Debian 12 "Bookworm" and shipping both Enlightenment 16 and 27 desktop environments. Uses OpenRC init system (optional during installation, for users who prefer not to use systemd). Available free in both 32-bit and 64-bit editions, one of the very few remaining distributions maintaining genuine 32-bit support. The "Retrowave" branding pairs a Synthwave aesthetic with an explicit e-waste reduction mission for 32-bit hardware.
• Source: 9to5Linux

Manjaro Governance Crisis: Seventh Week, Silence Persists

• The Manjaro governance standoff continues into its seventh week. No new statements from either Philip Müller or the 19 manifesto signatories. No fork announced. No nonprofit formation timeline proposed. Technical operations continue: stable and unstable updates ship normally. The complete absence of communication since March 19 suggests the crisis has settled into indefinite paralysis. The manifesto's Stage 3 option (fork or mass departure) remains unexercised.
• Source: It's FOSS