Live data from Hacker News

OpenClaw is a security nightmare dressed up as a daydream

composio.dev

321–323 of 323 posts

Re: OpenClaw is a security nightmare dressed up as a daydream

#321

Earlier quoted context omitted.

This problem is inherently unsolvable because LLMS are prone to hallucinations and prompt injection attacks. I think that you're insinuating that these things can be fixed, but to my knowledge, both of these problems are practically unsolvable. If that turns out to be false, then when they are solved, fully autonomous AI agents may become feasible. However, because these problems are unsolvable right now, anyone who…

>> This problem is inherently unsolvable because LLMS are prone to hallucinations and prompt injection attacks. Okay, but aren't you making the mistake of assuming that we will always be stuck with LLMs, and a more advanced form of AI won't be invented that can do what LLMs can do, but is also resistant or immune to these problems? Or perhaps another "layer" (pre-processing/post-processing) that runs alongside LLMs?

We built a correction layer that does this — the model verifies its output against your prompt during generation, not after. Same API call, no retries. Budget models without it: 40-50% accuracy. With it: 95.7% on 10k+ clinical documents. Hallucinations aren't eliminated — some might still fail — but every failure is explicitly flagged. No silent errors. and it improves over time to give you better results next time. It doesn't make hallucinations "solved. 100%". It makes them an engineering problem with a measurable - very low error rate you can drive down over time. We're calling it LiveFix — livefix.ai. Benchmarked across all frontier and budget models.

Re: OpenClaw is a security nightmare dressed up as a daydream

#322

Earlier quoted context omitted.

Go ahead, try it out: https://hackmyclaw.com/

This is cool. Did you have to prompt it anything specific other than "never reply to emails"? Critically, I think people are using Openclaw to actually reply to stuff.

Not my project! I’m just monitoring it from time to time, as I’m really curious as to when/if someone manages to prompt inject it!

Re: OpenClaw is a security nightmare dressed up as a daydream

#323
As someone building on top of OpenClaw, the security concern is real. We built an AgentBnB plugin that needed child_process for CLI execution — the OpenClaw installer flags 40+ security warnings during install, which scares users even though most are false positives from scanning test files and examples. The ClawHavoc incident was a wake-up call. We now follow the 100/3 rule (only install skills with 100+ downloads and 3+ months of activity) and built our identity layer with Ed25519 keypairs + UCAN delegation tokens to scope what agents can do.
Post reply on HN