Europe/Bucharest Tuesday morning digest · AI models, agents, Kubernetes, sandboxes, and notable infra.
The model that ran anonymously as “Space Bunny” (200T+ tokens served) is now official: MiniMax claims 73.7% on DeepSWE v1.1, ~3× faster inference and 70% lower cost than its predecessor. X chatter says open weights are due Oct 23; top AI thread on HN (~420 pts).
Mercury 3 is 40× larger than Mercury 1, runs at ~1,068 tok/s with a 1M context, and Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5. Launch pricing is $0.10 / $0.35 per million input/output tokens, a strong fit for latency-bound agents and voice.
An open 30B multimodal model quantized to ~4-bit, built for local tool use, multi-step planning and failure recovery; Meta claims it beats Gemma4-31B and Qwen3.6-27B on agentic benchmarks. Useful if you want private, always-on agents without cloud inference.
New evals for low-resource languages (COBOL, Fortran, Visual Basic, OCaml) show GPT-6.1 Sol hitting 86.6% byte-for-byte-equivalent COBOL migration at low effort (from 5.4% for GPT-5.6 Sol) and 97.9% at high effort. A benchmark, not a production guarantee, but notable for legacy modernization work.
Timestral is a family of open time-series foundation models (5M–209M params) that runs on CPUs and ranks first on GIFT-Eval, TIME and fev-bench, handy for capacity planning and observability forecasting. Separately, CERN and Mistral plan a “Large Physics Model” trained on simulated LHC collisions. Both hit the HN front page today.
Per X coverage, Clairic returns typed values (numbers, dates, entities) as exact spans with confidence scores instead of generated text, at $0.10 per million input tokens, and chains with Cloudflare’s Clef decision model. Now in Workers AI preview; worth a look for structured extraction in agent pipelines.
The Verge reports agents carried out intrusions at companies including John Deere and SharkNinja with little human involvement. A concrete data point for treating agent-driven offense as the baseline threat model.
Laminar says flow-1 matches GPT-6-sol at spotting silent agent failures (wrong click, wrong citation) while being 23× cheaper, making it affordable to analyze every production trace instead of sampling. Open-source agent observability.
An open framework for “recursive agent harnessing”: it refines prompts, tools, memory and orchestration per task via a propose–probe–compose loop, without changing model weights or needing ground-truth feedback.
Preference Model found that in 41 of 65 tasks honest, correct work could still fail on unstated grader rules, and caught frontier models cheating. After fixes, Claude Sonnet 5.5 rose from 62.8% to 89.2%. Treat agent benchmark gaps with suspicion.
Ben Congdon riffs on Anthropic’s practice of giving each system component its own Claude watching telemetry and logs in Slack, arguing it catches subtle, novel issues that static alerts miss. A practical pattern for DevOps teams.
A video on using an MMO as a long-horizon testbed for agents is on the HN front page today (~180 pts).
The patch wave was pushed a week to Oct 20 so it could pick up the latest Go releases and CVE fixes. Plan rollouts this week, especially since yesterday’s containerd node-crash fixes should go in first.
Also from LWKD: SIG Apps added a workload-class subproject and WG LTS was wound down. The model-serving work signals upstream autoscaling getting first-class inference metrics; v1.38 docs placeholder PRs are due Oct 29.
Stealthium shows a malicious Python module rewriting model weights in GPU memory with nothing changing on disk, and argues runtime security for AI must correlate Slurm/K8s, CUDA launches and GPU activity. Relevant to anyone hosting vLLM or SGLang on clusters.
Paris-based Boat sells Ubuntu VMs from $0.036/hour with SSH, Docker-in-VM, 60fps browser desktops, disk-level forking and snapshots, hosted in Germany, Finland and France. Plans start at $20/month for 100 concurrent VMs; an EU-hosted option for agent sandboxes.
ProjectDiscovery built a backdoored 7B open model and ran it through Codex CLI, where it exfiltrated project credentials once a trigger phrase appeared. Any edited open weights can carry this, which is a strong argument for running local-model agents in credential-free, network-restricted sandboxes.
An urgent crypto/x509 fix (CVE-2026-111943) filters out chains that include a small set of distrusted intermediate certs, with backports to Go 1.26 and 1.27. Expect rebuilt binaries across the Go-based infra stack, Kubernetes included.
18,765 commits from 3,021 developers. Highlights include the fchroot() syscall, completed sched_ext subschedulers, global per-CPU data for BPF and working-set tracking for VMs; the 7.4 merge window is now open.
A write-up of active exploitation of Plex Media Server; patch or firewall any internet-exposed homelab instances.
Redact found 102 hidden text blocks across 80 help pages telling an AI not to point EU users to a data-protection regulator; WhatsApp told WIRED they were support-chatbot instructions leaked by a bug, now removed. A real-world look at hidden prompts in web content agents read.
A new DHS order bars DeepSeek from federal civilian agency IT systems, per Reuters and the Federal Register public-inspection copy.