NIBVOK releases runtime AI security layer for OpenClaw agents

NIBVOK has published NIBVOK AI Security v0.1.2 in the ClawHub registry, a policy layer that makes allow, confirm, or deny decisions before an AI agent tool call runs. The release arrives as companies look for controls that stop harmful actions at the moment they are attempted, not after damage is done.

Why it matters: - AI agents can now run commands, write files, move money, and make other irreversible changes. - NIBVOK’s layer adds a decision point before each action, which is meant to block destructive behavior in real time. - The release targets a gap in many agent controls: matching a tool name instead of reading the command itself.

What happened: - NIBVOK published NIBVOK AI Security v0.1.2 in the ClawHub registry on Oct. 12, 2026. - The product is a policy-enforcement layer for OpenClaw agents. - The layer sits in an AI agent’s request path and decides whether to allow an action, ask a human first, or refuse it before execution. - The plugin is available as @nibvok-llc/nibvok-ai-security v0.1.2 in the ClawHub registry. - A product page is available here.

The details: - The layer is not a scanner, linter, or dashboard. - The system reads the command itself, which lets it confirm a force-push, deny a destructive wipe, and allow routine work to proceed. - NIBVOK says the product governs destructive deletes, secret material, writes outside the workspace, destructive git actions, database drops, any transaction at or above $500, changes to the governance layer itself, and outbound messages. - Outbound messages are always recorded. - The release includes a public test suite and a public incident log, INCIDENTS.md. - The incident log documents four policy bugs found during development and fixed in the classifier: over-broad write rules, reads conflated with writes, string-anchored secrets, and path traversal. - NIBVOK published a reproducible SHA-256 for the release artifact: 3c05f7d226f46a6c201a4613e1f9275ccc2b858f76588068ad1a154f970dc3bc. - The release artifact also shipped with a clean scan.

Between the lines: - The product is positioned as evidence-based security rather than a guarantee of perfect protection. - NIBVOK’s public bug log suggests the company is using its own development process as proof that the control layer was tested against failure modes it is meant to catch. - The emphasis on point-of-action enforcement reflects a broader shift in AI security from after-the-fact review to pre-execution control. - The release references high-profile AI failures and liability cases as examples of what can happen when an agent acts without a hard permission gate.

What’s next: - NIBVOK says the plugin is available now in the ClawHub registry. - The company says it will continue to position the product as a runtime control layer for OpenClaw agents. - NIBVOK directs users to its security page and media materials at the company’s announcement, media assets, and NIBVOK’s website.

The bottom line: - NIBVOK is betting that AI security needs to happen before an agent acts, not after a log file shows the damage.