Open Source Radar — October 24: VoxCPM2's 30-Language Open TTS
Three of today's five AI repos on GitHub's daily board are skill packs, not software — markdown methodologies an agent loads on demand. The other two are the hardware underneath: a voice model you can host yourself, and a runtime that decides what an agent is allowed to touch.
VoxCPM2 (2B params, Apache 2.0) — OpenBMB shipped the second version of its tokenizer-free speech model, and the spec sheet is unusually complete for an open TTS release: 2 billion parameters, 30 languages with no language tag needed, and 48kHz output reconstructed from a 16kHz reference clip, so you don't bolt on a separate upsampler. It is trained on more than 2 million hours of multilingual speech and runs in roughly 8 GB of VRAM, with a real-time factor near 0.30 on an RTX 4090. The feature that makes it a design tool rather than a demo is voice design — describe a gender, age, tone and pace in plain language and it invents a voice with no reference audio at all. LoRA fine-tuning works from as little as 5–10 minutes of audio. The commercial terms are the real headline: Apache 2.0, no revenue cap. OpenBMB's own model card puts last month's downloads at 915,673 (model card), and the caveats are honest — voice-design output varies between runs, so generate a few takes.
OpenShell (Rust, ~19,000 stars) — NVIDIA's sandboxed runtime for fleets of autonomous agents, now on a 0.1.x stable cadence after a long stretch of 0.0.x nightlies. The pitch has not moved, but the enforcement has thickened: filesystem access is confined by kernel controls (Landlock), dangerous syscalls are blocked by seccomp, and every outbound connection clears a network policy before it leaves the sandbox. Agents never hold real credentials — the runtime substitutes an opaque placeholder and resolves it only at an endpoint the policy already approved. What is new and worth the upgrade is the policy workflow: before a changed policy takes effect, OpenShell uses formal verification to flag the risky access it would newly grant, such as a new host reachable with credentials, and holds that change for a human. The honest caveat is maturity — 576 open issues, and the Windows path is still experimental under WSL 2. Reach for it when Claude Code or Codex is running unattended on a machine that also holds your SSH keys.
Claude-Red (Python, ~8,500 stars) — a library of 78 offensive-security skills across 23 categories, each a plain SKILL.md that primes the model with real methodology for one attack surface: 16 skills for web work, then authentication, Active Directory, wireless, shellcode, EDR evasion and exploit development. The design is the standard skills trick applied to a domain where it matters — a trigger in the conversation loads the relevant file, so an unused category costs no context, and a mention of SQL injection is enough to pull in the injection playbook instead of a generic answer. The author, SnailSploit, points to snailsploit.com for the research actually run with it. Two things to be clear about before you install it: every use case the README lists is authorized — engagements, bug bounty triage, CTFs, operator training — and the library enforces none of that, so the scope letter is your control. It also reads as the dark mirror of Cloudflare's audit skill that trended this week: attacker and defender methodology are now distributed the same way, as markdown.
Security teams: would you rather your agents carried a red-team skill pack you can read, or a general model you can't audit? Write to us via our contact page.
