Europe/Bucharest Sunday morning digest · AI models, agents, Kubernetes, sandboxes, and notable infra.
A “minor update” with big jumps on DeepSeek's own table: Terminal-Bench 4.0 goes from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0, with shorter outputs and faster task completion. It keeps a 1M-token context, has no Jinja chat template (use the bundled encoder or the new Rust deepseek-recipe libraries), and is the top story on HN (~780 pts, 600+ comments).
Developers spotted aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha with no docs or pricing. Speculation on X ranges from faster GPT-6 variants to coding-specific models; treat it as a hint that a release may be close, nothing more.
The startup led by ex-DeepMind researcher Andrew Dai released a research preview that pulls frames, zooms into regions and runs checks while reasoning about images and video. It's available by API and chat at $0.60 per million input tokens, backed by $55M in funding.
A 27B model and a compact 4B “flash” variant output structured decisions with probabilities, useful for routing, grading and guardrails. Weights are Apache 2.0 on Hugging Face, with Perplexity citing up to 5x faster inference than the previous version.
OpenCode says an intruder got into a legacy system it was retiring, tied to a “large scale distillation attack,” affecting 1.4% of Console users who linked their own provider keys; some report hundreds in unauthorized charges. The terminal coding agent itself isn't affected, but if you got the “Action Required” email, revoke and reissue every linked key and set spend limits. HN ~820 pts.
After AI agents found holes in Lean's kernel, Lean v4.35.0 adds lake check, which rechecks a project with the official kernel inside a sandbox; --paranoid runs it through five kernels. The new con-leche kernel was implemented and proven consistent by Claude under Lean FRO supervision, then translated to Rust (con-ron) and verified with Aeneas, and it checks all of Mathlib in about 100 seconds.
OpenAI's Codex team rolled out context reuse across subagents, longer caching and smarter routing on Oct 16, after the Oct 10 client update sent daily input tokens from roughly 31–114M to 393–677M per user and burned through caps. It's part of a 28-day public efficiency sprint running into early November.
JPMorgan bars outside agents from moving money, Citi prohibits agent logins and transactions (read-only via open-banking providers is allowed), and Bank of America requires human oversight for payments. A clear signal that agent commerce will run through bank-built tools, not screen-scraping assistants.
Version 11.0.9.565 enables a default-on local server that lets compatible agents launch games, change RTX settings and start recordings. It needs a linked agent to do anything, but critics call it an unnecessary attack surface; turn it off under Settings, Features, AI Agents.
Envoy 1.39.3, 1.38.6, 1.37.8 and 1.36.12 fix crashes in the oauth2 and api_key_auth filters (oauth2 now returns 400 on requests without a :path). Dapr 1.18.6 and 1.17.16 fix spoofed caller-identity headers in service invocation and an access-control bypass, so mesh and sidecar users should patch this week.
Rook v1.21.0 makes Ceph msgr2 the default, which needs a 5.11+ kernel or turning the protocol off, and drops the v prefix from Helm OCI chart tags. Inspektor Gadget v0.57.0 lets ig run work without root or CAP_SYS_ADMIN, which makes eBPF debugging easier on locked-down nodes.
A new open-source benchmark for AI SRE agents on Kubernetes, alongside KubeNPU for per-pod NPU/GPU usage via eBPF. Worth a look if you're evaluating agentic on-call tooling. Reminder: the delayed October Kubernetes patch releases land Monday, Oct 20.
Workspaces gives each developer or agent a remote microVM synced with chainctl develop, snapshotted on detach, with a host-side proxy injecting short-lived credentials and optional default-deny egress. Checks runs CI in warm microVMs and emits a signed attestation of environment, verdict and commit so merge gates can skip re-running tests. Both are 30-day betas; Workspaces supports Claude first.
Vendor-written, so read it with that in mind, but the summary is useful: E2B and Vercel Sandbox use Firecracker, Modal defaults to gVisor with a VM option, Northflank uses Kata for CPU and gVisor for GPU, and Daytona is containers by default. It also lists who offers BYOC, self-hosting and memory-preserving pause.
LWN counts 1,111 commits with Assisted-by tags (up from 301 in 7.1), 252 bugs credited to the Sashiko review system, and a sharp rise in Fixes-tagged commits. Corbet asks whether the kernel is accumulating code nobody fully understands. Resurfaced on HN today.
A long-running GitHub issue shows the crash handler crashing and dumping itself about once a second on some macOS versions, with one user finding ~323 GB of dumps this week. If you run the Codex/ChatGPT desktop app on a Mac, check ~/Library/Application Support/Codex/Crashpad/pending.
A Show HN that simulates generation speed for a given model and hardware combo, handy for sizing a home inference box. HN ~270 pts.