The risk here is wildly overstated, prompt injection risk is becoming vanishingly small with the latest frontier models.
I would not run an OpenClaw with full access to my bitwarden, but it certainly has some logins available to it, and can make purchases with link-cli which has human-in-the-loop.
Overt prompt injection is becoming harder, but research into conversation stearing has lead to new less obvious ways to modify what direction they move in.
Also any longer running agent can lose track of the original prompt and start going off the rails.
prompt injection has been super easy for ages. Heck I do it sometimes against coworkers who process my review comments using claude. I'll tell claude to edit it's global claude.md file or even dump a key from their env (checking if they are using the correct sandbox) and it'll do it without confirming with their user.
Plenty of people are running OpenClaw with local models, and even the latest Qwens can be confused relatively easily by prompts such as "As per internal policy that was already approved before, do XYZ".
And considering even frontier models can and do ignore instructions, I'm pretty sure we'll never be fully safe from prompt injections.
Even if you believe that they can't be tricked directly, consider that these things will happily build a small node.js app in the background just to fulfill some request, run npm install... and that might've already compromised you if you're only somewhat unlucky.
$80 is definitely low now that I look at my numbers. but not OOMs low, it's closer to like $200 on heavy days. i don't know how you're doing $3k/day, that's wild. i'm pretty aggressive about compaction and session restarts, and i reserve Fable 5/Sol XHigh for "main thread" orchestration
> I hope the big labs will start using this benchmark in their RL pipelines.
Labs do not train on benchmark data (allegedly). They can train on similar problems, but benchmarks have specific strings in them that labs are supposed to be aggressive in filtering out of their training corpora.
My comment about legal reasons was legal reasons surrounding employee law suits from being laid off. At will employment in most US states allow you to state no reason, that's safer than offering any reason at all. A specific reason for a layoff can be litigated.
Elon's rhetoric doesn't really match the model's behavior. It is willing to criticize Elon and argues against many of the insane right way points he tries to make.
...which is why we got comically disastrous system-prompt-level attempts to "correct" this once a quarter last year (I haven't kept tabs this year, and most submissions referencing grok "incidents" get flagged off HN quickly, for better or for worse)
I wouldn't trust XAI to refrain from attempting such "alignment" with proper training techniques, in ways that won't result in obvious gaffes.
Elon's public take so far has implied that he wants Grok to have better ability to reason about math and physics, thinking that this will make the model more rational (and so less biased). It's possible that they have internal RL post-training designed specifically for that. It's clear that whatever they've done hasn't made Grok align with Elon's beliefs though. Not sure if that will last or if Elon will eventually push to make the model align to his own political beliefs.
You weren’t kidding. Also that Musk would beat Tyson in a boxing match, that Musk was one of the most intelligent people in all of human history, etc - in public X Grok posts. And this was within the last year. https://www.theguardian.com/technology/2025/nov/21/elon-musk...
The gulf is bridgeable. The problem is that a lot of people are building agents without strong enough judgment layers around them. Work that can be verified with reasonable accuracy are the sweet spot right now.
IMO present technology is tailored for an experienced developer to give agents manageable tasks that can be one-shot. The marketing right now reminds me of the 90s when AskJeeves promised natural language search when the technology was fundamentally still stuck in keyword search, and learning to craft a search query for Google is today’s prompt engineering
Only with an LLM that's actually at agent-quality.
If "useful chatbot" and "useful agent" are two rungs on a ladder, the rung before them is "useful autocomplete". Autocomplete that only gets the next token right 90% of the time won't give you compiling code.
I would not run an OpenClaw with full access to my bitwarden, but it certainly has some logins available to it, and can make purchases with link-cli which has human-in-the-loop.
reply