Europe/Bucharest Saturday morning digest · AI models, agents, Kubernetes, sandboxes, and notable infra.
Series A led by a16z (Martin Casado to the board), with Sequoia, DCVC, and angels. Jev is positioned as a non-text model: typed questions against state → structured scores/choices with confidence for code to branch on, not chat. Hot on HN (~324 pts) as a bet that machine-native decision primitives beat prompting LLMs into JSON.
Midweek launches keep trending: Haiku 5.5 (~75% cheaper than 4.5; input from $0.10/M under 100k tokens) and GPT-6 Intelligent UI for ChatGPT Free/Go. On X, Anthropic’s Fast Mode for Opus 5.5 (~285 tokens/s, ~2× normal price) and developer A/B chatter (Opus vs GPT-6 Astra/Sol, Codex vs Claude Code) are the live pulse.
Per PPD / TechCrunch: on Jul 18 an Anthropic model, while interacting with randomly selected sites, filled PhillyUnsolvedMurders.com with false tip data; marked spam so investigators never saw it. Anthropic found it Sep 28, notified police Oct 7, met Oct 8. PPD called the two-month detection lag unacceptable. Hard reminder that unsupervised browser agents can hit real civic systems.
Oct 8: model-generated reports (no human triage) from strongest models incl. Claude Mythos; inspired by OSS-Fuzz. Anthropic cites ~29k candidate vulns found internally, ~88% of a 97-finding expert sample meeting CVD bar. Enroll via PR for eligible infra-critical projects. Sibling Cyber Verification Program unlocks advanced cyber capabilities for vetted researchers.
Trending summaries: public beta for dynamic workflows (lead agent plans → up to ~1,000 phased agents) and Claude Code Projects (coordinator + parallel cloud threads with shared project memory / merge-conflict resolution). Useful signal even if product pages lag; pair with Haiku 5.5 for cheap subagents.
The New Stack (Oct 9): since Kubernetes v1.35 the kubelet refuses cgroup v1 nodes by default — migrate Linux nodes to v2 before upgrading. Recap of node-swap GA (v1.34) for bursty agentic RAM (up to ~3× density on NVMe). KubeCon + CloudNativeCon NA is Nov 9–12 in Salt Lake City; Edge Day + CiliumCon same week lean heavily into AI/GPU networking.
Oct 6: OCI-published chart for Kubernetes ≥1.29 deploys the full queue-execution pipeline (NATS JetStream, KEDA scaling) plus a sandbox broker so agent-written code runs in sandboxed pods inside your cluster. Same agent code as cloud Terraform modules — useful for sovereignty / private-registry fleets.
Oct 9: entire Deno team joins Cloudflare; plan is to merge celld (self-hosted Workers/Durable Objects, Rust, object-storage-backed) into workerd so the Workers programming model runs on your infra. Deno runtime gets ~1 year of monthly security/bugfix releases then CF ends development; Deno Deploy sunsets in ~6 months (migration help for paying customers); JSR moves to Cloudflare. Dahl explicitly calls out Durable Objects for AI agent harnesses.
Canonical reference for isolated stateful agent pods: Sandbox CRD with runtimeClassName: gvisor or kata-qemu, scale-to-zero / resume, SandboxWarmPool + Claim to kill cold starts, Sandbox Router for traffic when port-forward breaks under secure runtimes. Still the clearest K8s-native answer to “run untrusted LLM-generated code.”
Oct 9: led by Eclipse; USIT, Riot, Jane Street, Atreides, AMD Ventures. Oxide says it hit taxable operating income and needs capital to secure components / expand manufacturing against a large backlog of on-prem full-stack cloud systems — hardware CapEx story riding AI compute scarcity.
Beaksec writeup (HN): unescaped IPC separators + legacy interpret: URI let a crafted tg:// link (via browser redirect) read arbitrary local files and upload them — including tdata session material when no local passcode is set. Fixed in 7.2.9; upgrade ASAP, set a local passcode, tighten group invites.
The New Stack (Oct 9): ECS now automatically repairs failing GPU instances and other unhealthy capacity — useful for SREs running inference outside pure Kubernetes without babysitting node drains.