2026-10-16
Top stories
- CGPM votes to stop leap seconds, making UTC continuous from 2027-05-20 UTC will stop taking leap seconds on 2027-05-20, and its gap from Earth rotation time UT1 may grow to as much as one hour.
- Meta releases Muse Spark Open, an 816B-parameter open-weight MoE model with a 1M-token context Meta's new open-weight model has 42B active parameters, is aimed at coding agents, and needs about 700GB of VRAM to run in NVFP4.
- Meta details how agents plan, build and verify its Forum app while engineers supervise Meta's Forum team required agent changes to pass builds, tests and UI checks before review, and gave agents tools to inspect the running app.
- PwnEscape: Linux checksum flaw CVE-2026-99999 enables guest-to-host KVM escape A Linux networking flaw lets guest-controlled checksum metadata rewrite a validated IPv4 header, leaking host memory and escaping KVM.
Security
- Common Crawl scan finds 30 sites sending prompt injection in HTTP headers aimed at AI agents A scan of 217 million crawled responses found 30 sites sending prompt injection headers to AI agents, and 22 of them were pranks.
- Calif publishes root exploit for a 37-year-old OpenBSD shared memory counter bug Calif chained a signed 16-bit shared memory counter from original BSD code with an out-of-bounds write to get root on OpenBSD.
- X.Org discloses 17 vulnerabilities in libX11, libXi, libXpm, libXrender and libXtst Out-of-bounds reads and writes and heap overflows in five X11 client libraries are fixed in libX11 1.8.14, libXpm 3.5.20 and other releases.
AI
- DeepMind uses Antigravity agent with ProtNLM to find antimicrobial peptide candidate DeepMind shows biologists using the Antigravity agent with ProtNLM and AlphaFold DB to find an antimicrobial peptide lead now in wet-lab tests.
- Reflection releases Beam, a 501B-parameter Apache 2.0 MoE model for coding and agents Reflection's first model is a 501B-parameter MoE with 23B active, a 1M-token context, Apache 2.0 terms and day-0 vLLM and SGLang support.
Agentic coding
- GNOME developer uses Claude Code to reverse engineer a scanner driver and releases hw-harness A GNOME developer used Claude Code with a Windows VM and usbmon traces to write a scanner driver and camera fixes, then released hw-harness.
- Dynata cuts feature development from three or four sprints to about one with AI-DLC in Kiro Dynata cut feature development from three or four sprints to about one by running AWS's AI-DLC method in Kiro with human approval at each phase.
- Claude Code team details layered AI code review that merges low-risk changes without human approval Anthropic lets low-risk Claude Code PRs merge without a human approval once AI, security and CI checks pass. Its revert share fell over 40%.
ML research
Developer tools
Linux and kernel
- Ubuntu 26.10 ships with Linux 7.3, GNOME 51 and a completed switch to Rust coreutils Canonical's 26.10 interim release adds RISC-V desktop images, amd64v3-optimized installs, and better TPM-backed full-disk encryption.
- Libreboot 26.10 makes EDK2 UEFI the default payload and adds 233 Chromebooks The free boot firmware now boots UKIs, systemd-boot and Windows through EDK2, and offers GRUB and SeaBIOS as EFI builds for compatibility.
- Fedora 45 misses its 2026-10-20 early target after a no-go decision No release candidate media was ready in time, so Fedora 45 moves to the 2026-10-27 or 2026-11-03 contingency dates, depending on blocker bugs.
- Linux kernel maintainers deprecate the Assisted-by tag for AI-assisted patches Kernel maintainers agreed to deprecate the Assisted-by tag, and Linus Torvalds rejected rules on which tools contributors may use.
Apple platforms
- Apple moves Developer ID certificates to a G2 sub-CA with one-year validity Apple now issues Developer ID certificates from a G2 sub-CA with one-year validity, and some signed pkg installers stop working on 2027-02-01.
- Apple runs Muse Spark Open on four Mac Studios with MLX over Thunderbolt RDMA Apple shows four M5 Ultra Mac Studios linked by Thunderbolt RDMA running the 816B Muse Spark Open model at about 36 tokens per second.
Infrastructure
- RPKI-valid routes reach 70.3% of routed prefixes, up from 54% a year earlier IPinfo measures 70.3% of routed prefixes as RPKI-valid in September 2026, up from 54% a year earlier, with China driving most of the gain.
- Neon launches Realtime public beta with live SQL queries running inside Postgres Neon Realtime enters public beta and keeps any deterministic SQL query, joins and aggregates included, live as Postgres data changes.
- PostgreSQL 19 slips to around 2026-11-19 after up to 10 features are reverted PostgreSQL 19 is now expected around 2026-11-19 after eight to ten features were reverted, partly due to LLM-assisted code review findings.
- StackGres reports open-source Citus at 10M QPS with 24 workers, against 50 data nodes for PlanetScale Neki A read-only pgbench test on Graviton4 put Citus with scaled-out query routers at 10M tps. The reproduction code is not yet published.
Engineering posts
- Meta handled record World Cup traffic with capacity buffers and no deployment freeze Meta says WhatsApp peaked above 30 million messages per second during the World Cup final while teams kept deploying without a code freeze.
- Spotify adds beam search and multi-item scoring to TensorRT LLM, cutting beam search cost up to 54% Spotify and NVIDIA closed TensorRT LLM feature gaps for LLM recommendations, replacing a split setup that used vLLM and SGLang side by side.