2026-11-01
Security
- Chainalysis discloses a breach in which an attacker used offensive agentic techniques to reach internal tools Chainalysis says an attacker using offensive agentic techniques reached internal tools and exposed customer names, emails, IDs and roles.
- Public PoC released for Android SMS of Death CVE-2026-58882, a zero-click crash of the phone process An eight-byte WAP Push SMS crashes com.android.phone and can loop reboots, and Android 12 and 13 devices did not get the October fix.
- Alpine Linux rebuilds its GitLab server and resets passwords after attempted CVE-2026-97229 exploitation Attackers targeted Alpine's GitLab with CVE-2026-97229 on 2026-10-20, and missing logs led the project to rebuild it and reset passwords.
AI
Agentic coding
- Anthropic adds a Claude Code rule on Opus 5.5 that steers it away from recreating specific existing games Claude Code on Opus 5.5 now refuses to build recreations of specific existing games, and developers report refusals on their own projects.
- Nolan Lawson benchmarks 12 JavaScript frameworks by the cost for Claude Code to build the same app Opus 5.5 built one Pokemon app five times per framework: Hono plus htmx cost least at $2.67 median, Angular most at $4.97 with 2 failures.
- Engineer details an OpenCode setup where GPT-6.1 Sol orchestrates and a self-hosted GLM-5.3-Flash writes code A practitioner routes about 70 to 75 percent of coding agent tokens to a self-hosted model on H200s while a frontier model plans and reviews.
ML research
- Paper finds on-policy distillation gains depend on where the teacher keeps reasoning paths open On-policy distillation gains come from matching where the teacher keeps reasoning paths open, measured by a new entropy shape correlation.
- Study of 25 open-weight models finds chain-of-thought reliance rises with scale and not with post-training Across 25 open-weight models, answers follow edited chain-of-thought steps more as models grow, and post-training does not change that.
- MTA-OPSD lets each token choose its own teacher in on-policy self-distillation MTA-OPSD lets each token pick its own teacher from a pool during on-policy self-distillation, gaining up to 5.9 points on Qwen3 math.
- numoracle replaces fixed atol and rtol with derived per-element bounds for testing low-precision GPU kernels A test oracle for Triton kernels from Liger-Kernel, Unsloth and vLLM caught 116 of 124 mutants with no false alarms on correct runs.
Developer tools
- KDE developer Eike Hein introduces Kwi, a GPU-accelerated terminal built on libghostty-vt KDE developer Eike Hein starts Kwi, a Qt Quick terminal on libghostty-vt that he measures at 3x to 35x faster than Konsole.
- chezmoi 2.74.0 removes the deprecated remove command and marks several integrations stable chezmoi 2.74.0 deletes the deprecated remove command and promotes ssh, docker, gopass and 1Password Connect support to stable.
Languages and runtimes
Linux and kernel
- Linux 7.4 merges OverlayFS fixes for bugs found by Chris Mason's AI review, several reachable by unprivileged users OverlayFS fixes for an info leak, use-after-free and NULL dereference bugs found by AI code review are headed to stable kernels.
- Linux 7.2.15 is the final 7.2 stable release, ending the series earlier than usual Greg Kroah-Hartman ends the 7.2 stable series early with 7.2.15 and tells users to move to 7.3.y to cut the number of live branches.
Apple platforms
Engineering posts
- Radim Marek measures DROP TABLE on large PostgreSQL tables: 16.1 s on ext4 versus 80 ms on XFS for 1 TB cold Dropping a huge Postgres table writes about 80 KB of WAL, but unlinking its files stalls replay on replicas, worst on fragmented ext4.
- Awide's Polar Proxy offloads reads inside open PostgreSQL transactions to replicas Awide's Polar Proxy offloads read committed SELECTs inside open PostgreSQL transactions to replicas using XID import and LSN waits.