security.

Move API keys from command-line arguments to a chmod 600 env file, hiding them from ps output.

Web-based API key management with masked display, inline editing, one-tap provider testing, and live process.env updates.

CVE-2026-25253 let one crafted link take over an OpenClaw gateway, and The Register counted more than 135,000 instances reachable from the internet. What happened, what was exposed, and what to do.

OpenClaw's history from sources: Clawdbot in late 2025, two renames in January 2026, CVE-2026-25253, 135,000 instances exposed, its creator's move to OpenAI and the OpenClaw Foundation of July 2026.

Securing an AI agent you host yourself: threat model, sandbox tiers, where credentials belong, Tailscale or SSH instead of a public port, logging, incident response.

Running an AI agent on your own server under GDPR: lawful basis, where data lives, LLM providers as sub-processors, DPIA triggers, the 72-hour breach notice.

vLLM shipped CVE-2026-48746 (auth bypass, CVSS 9.1) plus trust_remote_code RCEs CVE-2026-4944 and CVE-2026-5817. Upgrade to 0.22.1+ now and lock these settings.

ChromaDB (CVE-2026-45830 to -45833) and Weaviate (CVE-2026-59093) shipped multi-tenant authorization-bypass CVEs in 2026. The shared pattern, and the fix.

MLflow shipped three authorization and secrets CVEs in 2026: CVE-2026-2651 (critical), CVE-2026-8147, and CVE-2026-4035. What each one breaks and how to patch.

Ollama CVE-2026-5757 leaks heap memory to unauthenticated callers; Triton's DALI backend (CVE-2026-24213/24214/24264) has integer-overflow and OOB bugs.

CISA added nine Langflow, n8n and LiteLLM CVEs to its exploited list in 2026. The versions to run, the three bug patterns, and what to lock down.

CVE-2026-103663 lets an unauthenticated caller of Ollama's /api/pull write files outside the model store and get code run as root. Upgrade to 0.35.0 and close port 11434.

NVD published 19 Open WebUI CVEs on 12 October 2026, all fixed in 0.11.4: session token theft from any website, terminal privilege escalation and stored XSS.

vLLM's autumn 2026 CVEs: audio and video inputs that crash or overflow the server, a trust_remote_code bypass, and disaggregated-serving DoS bugs. Run 0.30.0 or later.

Third-party add-ons around OpenClaw and Hermes Agent picked up CVEs in October 2026: an unauthenticated prompt-injection port, a skill that deletes your home directory, a config injection that runs shell commands.

Agent from Nous Research (MIT, Python). Runs shell commands on the host by default, with an approval step for dangerous ones; Docker and cloud sandboxes are options. Works with many LLM providers.