2026-10-19
Top stories
- Linux 7.3 released with a record 18,765 commits, opening the 7.4 merge window Torvalds shipped Linux 7.3 on schedule, adding fchroot(), completed sched_ext subschedulers, and BPF global per-CPU data.
- Plex2Shell RCE chain in Plex Media Server is widely exploited after a public PoC on 2026-10-16 Four chained Plex Media Server CVEs give unauthenticated code execution, and attackers are using them to install cryptominers and proxyware.
Security
- Broadcom WPS parser bug gives unauthenticated RCE over WiFi on unpatched ISP routers, with no CVE A WPS EAP reassembly overflow in Broadcom's 802.11 SDK lets attackers in WiFi range leak memory and get root on cable routers.
- NorForge finds 20+ exposed Cairn panels running autonomous AI agent intrusions against web shops Open Cairn dashboards show criminals feeding AI agents target lists, with one Chinese-language cluster running over 1,100 attack projects.
Outages
AI
- Mistral releases Timestral, a family of CPU-friendly time-series forecasting models from 5M to 209M parameters Mistral released four time-series forecasting models from 5M to 209M parameters that run on CPU and take past and known-future covariates.
- MiniMax launches M3.1 Flash for agent and coding work at $0.15 and $0.60 per million tokens MiniMax launched M3.1 Flash for agentic and coding work at $0.15 input and $0.60 output per million tokens, with open weights promised later.
- DHS orders federal civilian agencies to remove DeepSeek and third-party products built on it A DHS exclusion order bars DeepSeek models and tools built on them from civilian federal IT and purchasing, with a 60-day exception window.
- Inception releases Mercury 3, a diffusion LLM claiming 1,068 tokens per second and a 1M-token context Inception's largest text-diffusion model costs $0.20 and $0.70 per million tokens and is available through OpenRouter, Baseten and Vercel.
Agentic coding
- vLLM ships agent-built CuTe DSL decode path for Kimi K3 and proposes a fast-track review model NVIDIA and vLLM used coding agents to write fused GB300 decode kernels for Kimi K3 and now bound their review cost with tests and isolation.
- Cloudflare cut its AI code reviewer's cost per review 68% with prompt cache fixes and fewer LLM calls Cloudflare cut its AI code reviewer from $2.43 to $0.77 per review by fixing prompt cache prefixes and dropping wasted LLM calls.
- Meta makes data pipeline agents more reliable with a thin deterministic orchestrator and 14 subagents Meta routes 20+ data pipeline workflows to 14 specialist subagents via a deterministic 1,300-line orchestrator that does no domain work.
- Google Research open-sources Harness2, which refines agent harnesses per task without changing weights Harness2 edits prompts, tools, memory, skills and subagents around a fixed model and reports up to 11.8-point gains on professional benchmarks.
ML research
Linux and kernel
Infrastructure
- Let's Encrypt adds two remote validation perspectives, with staging on 2026-10-27 and production on 2026-11-10 ACME domain validation will run from one primary and six remote vantage points, up from four, to meet a CA/B Forum rule taking effect 2026-12-15.
- PostgreSQL 19 GA moves to 2026-11-19, with RC2 on 2026-11-12, after a cycle of late feature reverts PostgreSQL 19 RC1 shipped on 2026-10-15, and the project now targets RC2 on November 12 and general availability on November 19.
- Meta details Prometheus, its Ohio GPU supercluster, with 85% pretraining goodput and in-rack spare GPUs Meta runs 64 of 72 GPUs per NVL72 rack and keeps eight as spares, so a GPU failure restarts from checkpoint without waiting on repair.