OpenClaw picked up 247,000 GitHub stars in roughly 60 days. That kind of growth is impressive, and also a little alarming when you look at what came with it: 40,000+ exposed instances, credential leakage, documented data exfiltration, and a supply chain attack called ClawHavoc that poisoned around 20% of the skill registry. A critical remote code execution vulnerability rounded things out.
Two separate projects have launched to address this, coming at the problem from different angles.
<h2>Jentic Mini: a permission firewall for agent API access</h2>
Dublin startup Jentic built Jentic Mini after 18 months of enterprise work on the same problem. It’s a free, open source, self-hosted layer that sits between AI agents and the APIs they call.
The core idea is straightforward: agents shouldn’t hold credentials directly. Jentic Mini stores them centrally and brokers access, so an agent can draft an email but not send one, or read a calendar but not modify it. When something goes wrong, there’s a single killswitch to cut all agent data access immediately rather than hunting down credentials scattered across a dozen integrations.
The tool ships with a catalog of 10,000+ APIs and works with OpenClaw, NemoClaw, and other general-purpose agents. Jentic is already a verified connector in Anthropic’s Claude.
Jentic CEO Sean Blanchfield has also been making a broader argument: as agents get better at building their own tools on demand, the case for buying pre-packaged SaaS software weakens. That’s a separate conversation, but it’s the context in which Jentic Mini exists.
<h2>DefenseClaw: governance at the execution layer</h2>
Cisco’s VP of [role] took a different approach, starting from a personal one: he runs OpenClaw at home. That hands-on experience with the same vulnerabilities everyone else hit led to DefenseClaw, an open source governance layer built on top of NVIDIA’s OpenShell sandbox.
DefenseClaw focuses on three things:
<ul>
<li>Pre-installation scanning of skills, tools, and plugins using five integrated scanners before anything runs</li>
<li>Runtime threat detection via content inspection at the agent execution loop</li>
<li>Hard-enforcement block/allow lists with sub-2-second revocation</li>
</ul>
All telemetry streams into Splunk out of the box, which means every agent action is observable from day one rather than reconstructed after an incident. The project launched March 27, 2026 on GitHub.
<h2>The underlying problem</h2>
Both projects are responding to the same gap: AI agents that move fast and touch real systems, with security tooling that hasn’t kept up. OpenClaw’s growth exposed that gap quickly and publicly. Whether you care more about credential management and API permissions (Jentic Mini’s focus) or pre-execution scanning and runtime monitoring (DefenseClaw’s focus), the problem is real enough that two independent teams built tools for it within the same window.
Neither project is a complete solution on its own. But the fact that they exist at all suggests the ecosystem is starting to take agent security seriously rather than treating it as someone else’s problem.

Collections

OpenClaw, an AI agent framework that gained 247,000 GitHub stars in ~60 days, brought serious security problems: 40,000+ exposed instances, credential leakage, data exfiltration, a supply chain attack (ClawHavoc) affecting ~20% of the skill registry, and a critical RCE vulnerability. Two open source projects have emerged to address these gaps. Jentic Mini, from Dublin startup Jentic, is a self-hosted permission firewall that brokers API credentials centrally so agents never hold them directly, with a killswitch for immediate revocation and support for 10,000+ APIs. DefenseClaw, built by a Cisco VP, layers governance on top of NVIDIA's OpenShell sandbox with pre-installation skill scanning, runtime threat detection, hard-enforcement block/allow lists with sub-2-second revocation, and out-of-the-box Splunk telemetry. Together they represent the ecosystem beginning to treat agent security as a first-class concern.

OpenClaw's security problems are real. Here's what two teams are building to fix them

I think at this point vibe coders are fully aware they are running through a 🌽 corn field naked backwards as soon as they deploy production.