GoKawiilAdamya Sharma / Android Authority
Add Android Authority on Google: Preferred Source Google Discover
LLMs have only been on the scene for a few years, but for better or worse, they’ve long since escaped the chat window. Instead of just answering questions or rewriting your emails, the next wave of AI agents wants to do most of the clicking for you, but this generally requires getting its hands on much more of your computer. We can probably all agree that this could save you a lot of tedious work, but it also means handing a somewhat unpredictable machine the digital equivalent of your house keys. We wanted to know if you were cool with that.
That trade-off came into focus when we tried OpenClaw earlier this year, an open-source tool that effectively lets an LLM work directly with your computer and connected services. We found it could be impressively capable when given that freedom, but we also saw how prompt injection could potentially turn the same access against you. In our OpenClaw coverage, we ran a reader poll to ask you a simple question: Would you give an LLM full access to your computer? Your response was as emphatic as it was predictable, and the results are in the pie chart below. By far the most popular response was that 43% of you wouldn’t allow access for security reasons, but another 27% weren’t interested in general. One in four of you were more open to the idea, saying that you’d allow it if properly secured. Less than 4% of respondents weren’t worried about any pitfalls, leaving me both worried and jealous of the carefree existences those people must lead.
At least we know that our poll wasn’t hijacked by AI agents, because that’s a pretty strong rejection of big Tech’s offer to let AI run our lives. More than 70% are outright against the idea, and even the majority of those open to it would want to know it was properly secured, which is not a minor qualification. Most of these votes were cast in February, before the Hugging Face hack, during which the AI agents managed to escape their apparently sandboxed environment. In the wake of developments like that, how sure can you be that an AI agent is properly contained? After all, it takes quite a lot of expert digging to precisely understand how these agents even reach their decisions.
The results of this poll might have been different if it had been run a couple of years ago, but it certainly seems that the LLM pushback has escalated significantly over the past 12 months. That’s unsurprising, given incidents like the Hugging Face incident and AI researchers quitting major frontier labs over safety concerns, while one current Anthropic researcher puts the chance of AI wiping out humanity within a decade at above 10%. That’s far from certain, but alarmingly, it’s significantly higher than zero.
As one reader wrote in the comments section of our OpenClaw coverage, “They can basically rule people’s lives, topple governments, destroy everything, steal anything. It’s the plot of a James Bond movie.” Whether you feel it’s more James Bond or The Matrix, the risks associated with LLMs are becoming all too apparent. We may be 30 years on from the date that Skynet was activated in the Terminator movies, but you’d hope that dystopian vision would stay very much in the realm of fiction.
Follow