What annoys me most about OpenClaw after trying it for a few weeks is that it cosplays security so incredibly hard, it actually regularly breaks my (very basic) setup via introducing yet another vibe coded, poorly conceptualized authentication/authorization/permission layer, and at the same time does absolutely nothing to convince me that any of this is actually protecting me of anything.

Maybe this idea is lost on 10^x vibecoders, but complexity almost always comes at a cost to security, so just throwing more "security mechanisms" onto a hot vibe-coded mess do not somehow magically make the project secure.

Agreed! Made my own OpenClaw variant based on many of the same principles. It takes Simon Willies lethal trifecta and implements it to an OpenClaw like architecture.

https://www.tri-onyx.com/

I love how despite all this, the author still uses the language:

> We’re simply not there yet to let the agents run loose

As if there aren’t fundamental properties that would need to change to ever become secure.

Related: https://news.ycombinator.com/item?id=47475997

Have you tried NemoClaw? Not endorsing it; just enquiring. Nvidia is claiming to provide guardrails with that.

One thing I'd like to critisize - although I can agree that skill security is a real problem, but the solution is not to restrict yourself from using them, but to rely on the community: reviews, likes/dislikes, maybe having the skills curated. We need some trust signals. Also, since markdown files are auditable by design - your agent might actually verify them before running - provided you're using something like GPT-5.4 on high reasoning.

This read like an AI generated piece and seems to be an advertisement for their product.

Should have said this was a fear to promote a b2b sass "TrustClaw"

Didn't Nvidia create the safer version basically?

Every LLM evangelist seems to forget that there is a reason why LLMs work so well for coding. It's because there were and are preexisting non-LLM validation tools for coding. The slop doesn't make it past linters, compilers, cone analysis and other tools, and then there is a second barrier in the form of code review. And even will these guardrails LLMs often produce substandard output.

Buying a ticket, writing an email, setting calendars or fiddling with files on the drive etc. have none of these guardrails. LLMs can and will simply oneshot the slop into a real system, without neither computer nor human validation.