A 753B open-weights model dropped under MIT, GitLab shipped secret-scan GA, and malware authors are now stuffing CBRN keywords into comments specifically to blind the AI tools scanning for them.

1. Web-Based Remote Installation for Fedora Linux: Here’s What We’re Building — Fedora Magazine · Jun 17
   The Anaconda installer team is building browser-based remote installation for Fedora Linux, replacing VNC and RDP with a Cockpit-served HTTPS interface that lets you point any browser at a machine mid-install and control it with a PIN. The Web UI already shipped in Fedora 42 Workstation and was extended to all Live spins in Fedora 43; the remote access piece is currently a proof-of-concept (PR rhinstaller/anaconda-webui#1274) with PIN auth hardcoded to 1234, no TLS yet, and no single-connection enforcement. The design calls for self-signed certs (browser warning expected), one concurrent session only, and Cockpit config isolated under /etc/anaconda/cockpit/ so it doesn’t bleed into the installed system. The practical upside over VNC is real: a remote browser sends far less data than pixel-streaming a full desktop, and no client software is needed. The team is soliciting feedback now on one open question in particular – whether to ship a lightweight headless ISO with remote installation on by default, dropping the bundled local browser entirely.
2. Ship AI with Laravel: Give Your AI Agent Live Web Search — Laravel News · Jun 18
   This Laravel News tutorial (8 minutes) shows how to add live web search to a Laravel AI support agent using the SDK’s built-in WebSearch tool, capped at five results and locked to an allowlist of shipping carrier domains. Without the allowlist the agent can browse anywhere and return arbitrary content to customers, so that’s the config you don’t skip. The practical demo has the agent checking FedEx’s live status page alongside order data instead of guessing from stale training data. WebFetch (for fetching a specific URL) only works on Anthropic and Gemini right now, so the series stays with WebSearch on OpenAI. The agent now has five tools total – order lookup, customer history, knowledge base, document search, and live web – with a production-infrastructure episode covering queuing, provider failover, and rate-limit middleware coming next.
3. GitLab 19.1 released — GitLab · Jun 18
   GitLab 19.1 shipped on June 18, 2026 with a handful of features worth scanning if you run Ultimate tier or manage AI tooling. The headline addition is GA secret false-positive detection via GitLab Duo – it automatically analyses critical and high severity findings after each scan and attaches a confidence score, aiming to cut the alert fatigue that makes secret detection results hard to trust. More practically useful for most teams: secret detection on feature branches now scans every commit from the branch’s divergence point with the default branch, not just the latest push, so credentials leaked in earlier commits no longer slip through to shared branches undetected. Admins get two new governance controls – a symmetrical “always on” lock for GitLab Duo (matching the existing always-off setting) and per-tool approval policies for Duo agents, letting you gate write or destructive operations with an inline human-approval prompt rather than trusting the agent to self-limit. Code Review Flow also adds GPT-5.2 and GPT-5.3 Codex as model options alongside Claude, which matters if your org has procurement constraints on Anthropic; both models benchmarked comparably to Claude Sonnet 4.6 on GitLab’s internal dataset. Nineteen compliance framework templates – including ISO 27001:2022, SOC 2, and FedRAMP – are now available in beta to reduce the manual grind of building controls from scratch.
4. Cloudflare: Build your own vulnerability harness — Cloudflare Blog · Jun 18
   Cloudflare’s second Project Glasswing post details the architecture of their model-agnostic vulnerability scanning pipeline, covering 128 repos in six languages with no per-language tuning. The core design splits discovery and validation across two different frontier models – one for the Vulnerability Discovery Harness, a separate one for the Vulnerability Validation System – so Model B is independently stress-testing Model A’s findings rather than rubber-stamping them. Key practical findings: single-run coverage catches roughly half the bugs you’d surface across multiple runs and skews toward simpler issues; each agent is kept below 25% context window usage to prevent hallucination; state is persisted to SQLite keyed by (run_id, repo, stage) so a crash costs only the in-flight task. The post is genuinely operational rather than a benchmark showcase – it includes specific advice on transient API errors returning as 200 OK text, nested container sandbox flags, and when not to build deduplication agents yet. Worth reading if you’re designing any multi-agent pipeline that needs to survive across runs at scale, not just security tooling.
5. GLM-5.2: 753B MIT-licensed text-only open weights model from Z.ai — Simon Willison · Jun 17
   Z.ai released GLM-5.2 on June 16th under an MIT license – a 753B parameter, 1.51TB MoE model with 40 active parameters and a 1 million token context window, up from 200k in GLM-5.1. It’s text-only (no vision), but that doesn’t seem to hurt it much: Artificial Analysis ranks it first among open weights models on their Intelligence Index v4.1 at a score of 51, ahead of MiniMax-M3 and DeepSeek V4 Pro at 44. It also sits 2nd on the Code Arena WebDev leaderboard behind Claude Fable 5. The catch is token verbosity – it burns through 43k output tokens per benchmark task, well above GLM-5.1’s 26k and most peers in the 24-37k range, which matters for cost. At $1.40/$4.40 per million tokens on OpenRouter, it’s substantially cheaper than GPT-5.5 ($5/$30) or Claude Opus 4.5-4.8 ($5/$25), but the extra token usage narrows that gap in practice.

---

// In other news

GLM-5.2 weighs 1.51TB and someone’s already hiding weapons text in comments to blind the scanners — good week to read slowly. Back Monday.