Akamai opens OpenAI-compatible inference for 12 open models
Beta serving runs gpt-oss-120b up to a 2.8T model behind one base URL; a WASM agent sandbox ships too, as critical fixes hit MCP Toolbox, MLflow and Tailscale.
Copy markdown
12 open models behind one OpenAI-compatible URL
Akamai Inference (beta) serves 12 open-weight models from 8B to 2.8T params - gpt-oss-120b, NVIDIA Nemotron 3 Ultra, Google Gemma 4, plus RAG embedding models - over OpenAI-compatible APIs on serverless, pay-per-use. Point your base_url at Akamai's edge (4,350+ locations) and existing code runs unchanged. Pricing isn't public yet.
A WASM jail for agent-written code
The companion Agent Sandbox (beta) runs your agent on Akamai Functions and executes the code it generates in a separate WebAssembly-isolated environment that caps compute, memory, outbound connections and syscalls. It's an open-source component of the Spin framework - a drop-in answer to 'let the model run this' without handing it your box.
Patch Google's MCP Toolbox now - CVE-2026-93328 (9.3)
Google Cloud's MCP Toolbox for Databases (1.2.0-1.4.0) ships an unauthenticated path-traversal in its cloud-storage source: a crafted tool call lets a remote attacker read host files into an external GCS bucket and overwrite local files, with no auth or user interaction. If your agents touch it, upgrade past 1.4.0 (fix in PR #3454).
MLflow RCE on any no-auth server (CVSS 9.8)
A scorer-deserialization flaw (GHSA-26p8-2jq9-3vq9, tracked as CVE-2026-116692) lets an unauthenticated user upload a Python module and run it as the MLflow server process on default servers - versions 3.12.0rc0 through 3.16.x. Upgrade to 3.17.0, and never put MLflow on a network without auth.
Tailscale SSH 'Tailworm': mitigate before the patch
Researcher Erica Windisch disclosed 11 SSH findings (no CVE yet) where a foothold on one tailnet device escalates to root on others when Tailscale SSH port forwarding is on. Tailscale said Oct 27 it built fixes and is targeting a release within ~24h. Set TS_SSH_DISABLE_FORWARDING=1 now - essential on self-hosted Headscale, which can't toggle forwarding per rule.