Windows 11 26H2 arrives as an enablement package, with memory tooling and admin lock-downs inside
Microsoft published documentation for Windows 11 version 26H2 on 30 September. It confirms the annual feature update ships as a small enablement package on the same servicing branch as 25H2 and 24H2, and switches on features previously held behind temporary commercial controls.
The page is not a launch announcement. It is the IT pro "what's new" note on Microsoft Learn, dated 30 September. It describes an update that most admins will recognise before they install it.
26H2, 25H2 and 24H2 share one servicing branch. Microsoft says many features listed there have already arrived through monthly quality updates. The annual release "enables selected features and resets the support lifecycle for the new release", according to the documentation.
That framing matters for anyone planning a fleet migration. Eligible devices move to 26H2 through an enablement package, not a full operating system upgrade. Microsoft says this reduces validation effort because the supported versions share the same underlying platform. Management, security and update experiences stay consistent across those releases. The company also repeats its continuous innovation line. Some features organisations associate with 26H2 may already be on devices before the feature update is installed, and the update activates and consolidates them.
What actually switches on
Three items in the document stand out for commercial deployments. Windows settings backup becomes enabled by default for eligible commercial devices. It preserves supported Windows settings and Microsoft Store app lists for recovery and device replacement. Existing administrator-configured policies continue to be honoured. App-specific taskbar actions also turn on by default for commercial devices. Several File Explorer improvements that sat under temporary commercial controls become default too, aimed at navigation and usability.
Security features are present but off until someone turns them on. Administrator protection lets administrative tasks run with just-in-time privileges instead of standing administrator rights. It also introduces profile separation to harden Windows against elevation-of-privilege attacks. Microsoft says it is off by default and can be enabled through Intune or Group Policy. Sysmon, the system event monitor used for threat detection, is now built into Windows. It supports custom configuration files to filter events and writes captured events to the Windows Event Log so security tools can consume them. It is also off by default.
Smart App Control can now be switched on or off without a clean installation of Windows. When enabled, it blocks untrusted or potentially harmful apps. The document also notes that Windows supports plugin credential managers for passkeys. The excerpt available cuts off mid-sentence and does not describe the mechanism further.
The feature is off by default. You can enable it using Microsoft Intune or Group Policy.
That sentence, about administrator protection, is the pattern across the whole page. Capability delivered, control left with the administrator.
Why the memory question keeps coming back
None of this is a context window story in the AI sense, and it would be wrong to dress it up as one. But the two conversations keep colliding in practice. Agents running on developer machines need isolated environments. The tooling built to provide them has its own memory and permission problems.
Martin Tapia, writing on 30 September, described testing Orca and the VSCode Agents window as agent development environments. He returned to tmux shortly after. His central complaint is about reach, not speed. The Orca CLI, he writes, is available to agents. It lets any agent list any shell open in the IDE and send arbitrary commands to it. He asked for confirmation and says the CTO replied that this is intended behaviour.
He reports the same class of problem in the VSCode Agents window. When you SSH into an agent sandbox, he writes, agents can climb back to your VSCode instance. They can install extensions or open arbitrary links.
He also flags automatic SSH agent forwarding by default. Sandboxed agents can then SSH as the user into other machines with the same access rights. He notes a large transfer to the sandbox from VSCode remote that runs to gigabytes of disk usage. The model picker, he adds, is bugged at the time of writing. It does not let users select models present on the agents in the sandbox.
These are one developer's observations, published on his own blog, and should be read as such. They do line up with the direction Microsoft's own documentation takes. Administrator protection exists precisely because standing privileges are dangerous, and it ships switched off.
The plumbing underneath
Memory safety in extensions is a separate thread, and a more technical one. Philip Dubé wrote on 30 September on the ClickHouse blog about bringing C++ into Postgres extensions, and the mismatch at the centre of it. Postgres allocates through palloc and pfree inside MemoryContext arenas. It handles failed allocation by raising an error, using PG_TRY, PG_CATCH and PG_FINALLY macros built on setjmp and longjmp. C++ uses new and delete, invokes destructors, and raises std::bad_alloc.
The two do not interact well. As Dubé puts it, setjmp and longjmp skip C++ cleanup. Jumping past non-trivial destructors for automatic objects is undefined behaviour rather than merely a leak. Exceptions bypass PG_CATCH and PG_FINALLY. An uncaught exception aborts the process. In a background worker, that forces the postmaster to restart everything in case shared memory has been corrupted.
Each extension he describes took a different route. For pg_clickhouse the team removed C++ entirely. They replaced clickhouse-cpp with a new C library, clickhouse-c, wrapped in pg-clickhouse-c. The Postgres logic sits in one place while clickhouse-c stays general purpose. The library is transport-agnostic and supports Native format outside the Native protocol. The HTTP driver now shares decoding and encoding logic with the binary driver. For HTTP they use libcurl.
For pg_re2, C++ is confined to re2_wrapper.cpp. All allocations sit inside C++ try and catch blocks, with no calls into Postgres C code. pg_chdb reuses pg-clickhouse-c for decoding and encoding native blocks. It moved libchdb into a separate helper executable, so its C++ runtime, threads, allocations and failures live outside a Postgres backend or managed background worker. The backend forks and execs the helper. The helper stays a descendant of postmaster without being managed as a background worker.
A high severity OpenSSL bug
Elsewhere in the stack, OpenSSL published a security advisory on 29 September covering a high severity DTLS flaw, CVE-2026-84782. The retransmission logic does not correctly handle a handshake message write suspended part-way through. A retransmitted message can therefore be read past the message buffer. The retransmission also overwrites internal state the suspended write needs to resume.
The advisory says the retransmitted message can disclose heap memory to the peer as plaintext handshake data. It can also cause a crash and denial of service when the read reaches an unmapped memory region. It is classified CWE-125, out-of-bounds read. OpenSSL 4.0, 3.6, 3.5, 3.4, 3.0, 1.1.1 and 1.0.2 are listed as vulnerable. Fixed versions are 4.0.3, 3.6.5, 3.5.9 and 3.4.8, with 3.0.23, 1.1.1zj and 1.0.2zs for premium support customers. Laurent Gaffie of secorizon.com reported the issue on 17 August 2026, and Ryan Hooper developed the fix. FIPS impact is listed as no, because the affected code sits outside the FIPS module boundary.
The same advisory covers a moderate use-after-free in the X.509 extension cache under concurrent use, CVE-2026-84783. The first concurrent use of a certificate by several threads may free cached extension data while another thread is still using it.
Sovereignty, and what it costs to leave
On 29 September, OpenSourceWatch reported that the Netherlands is building its own government Linux distribution and open-source office suite. The name is Digitaal Autonome Werkomgeving Overheid, or DAWO. The stack is in limited pilots, built by the Ministry of the Interior and Kingdom Relations with state IT organisations SSC-ICT, DICTU and DUO-ICT. Eight towns are running trials according to that report.
The outlet says developers tried openSUSE and Fedora before choosing NixOS. It cites declarative configuration and the Nix package model as reasons it suits fleet rollouts, reproducible desktops and rollback. France's La Suite and Germany's openDesk are being considered for the desktop suite. A possible home-brew of MijnBureau assembled from Nextcloud, Collabora, Element and OpenProject is also in play. The report also notes that the Swiss military and other agencies announced a move to Linux and openDesk earlier in September. It lists Austria, France, Germany and Ireland among earlier European shifts. Those claims are the outlet's own reporting and are attributed here to it.
For admins, the practical read is less dramatic than any of these headlines. A feature update that mostly flips switches. A set of security tools that stay dormant until configured. And a growing pile of reasons to check what your agent tooling can reach on your own machine.
Sources
8- 01What's new in Windows 11, version 26H2EN
- 02I tried Orca and VSCode Agents window: I went back to tmux shortly afterEN
- 03Memory safety for Postgres extensions in C/C++EN
- 04OpenSSL Security Advisory [29th September 2026]EN
- 05The Netherlands is preparing to dump Windows and Office for its own homebrew LinuxEN
- 06CoW - a stacking window manager for WaylandEN
- 07PCIe Part 2 - All About Memory: MMIO, DMA, TLPs, and more!EN
- 08The Ecology of Memory: Scent, Culture, and the BrainEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.