THE biggest election year in the history of the world is under way, and we have just had our first glimpse at how artificial intelligence is being used by the shadowy world of government-backed hackers. These groups could have significant impacts on democratic processes through hacking, disinformation or leaking sensitive information. In February, and published a blog in which they identified groups affiliated with Russia, China, Iran and North Korea and detailed how they were using AIs like OpenAI’s ChatGPT.
Good news, you might think. Crisis averted! But just as the veil has been slightly lifted, this new window onto chatbots’ illicit uses could already be closing due to increasingly capable private and open-source large language models (LLMs). And even if it stays open, it remains to be seen whether identifying these threats will have any impact on preventing real-world harms.
The blog itself wasn’t immediately alarming. One finding that Microsoft and OpenAI were quick to highlight is that there was no evidence of “significant attacks” orchestrated using their LLMs.
Advertisement
Instead, the companies produced a laundry list of smaller misdeeds, such as “LLM-informed reconnaissance”, which means asking chatbots for help explaining complex, valuable technologies or security systems, specifically used by a Russian military-linked “actor” looking for satellite and radar information in the context of the Ukraine war.
They also found “LLM-supported social engineering”, tricking people into giving away important information, used by a suspected North Korean outfit that was looking for help crafting convincing phishing campaigns – cyberattacks used to steal sensitive data like passwords. Programming tasks were also common, such as asking for help to develop computer viruses or finding ways to circumvent antivirus software.
According to Microsoft and OpenAI, this was all in line with their safety testing and predictions for how their systems might be misused, and they closed accounts associated with these groups. They also said they would continue monitoring their tools for further misuse. But there is a big issue here that didn’t come up: the rapid rise of open-source AIs that can be run without access to the internet.
One major reason why these companies were able to identify and catch these threats at all is because their technology, on the whole, is better than everyone else’s, so they have a large user base. For many people, and that includes state-backed cybercriminals, there are some tasks that only ChatGPT can do.
But other companies and models are rapidly catching up. For many programming tasks, open-source AIs that people can tinker with and run on their own computers can do the job. Companies like French start-up Mistral and Meta’s Llama models have proved surprisingly popular with people turned off by OpenAI’s closed ecosystem, which people can’t run on their own machines. And while running these models on local machines is a win for data privacy and experimentation, these models can be run offline, so are also far from the all-seeing eye of Microsoft’s cybersecurity experts.
While Microsoft and OpenAI said their models didn’t appear to have been used for significant attacks, the stakes are higher this year, when more than half the world’s population, including India, the US and UK, will have the chance to vote for new governments. Already, we have seen how AI can cause chaos for elections in the form of digital deepfakes, affecting democracies in Slovakia, the US and the .
There are some obvious ways to mitigate against deepfakes, such as creating certificates of origin for online media, but protecting against the text-based tasks that LLMs deal with, like a fake email written to sound official, is trickier. Social engineering scams and operations rely on the fallibility of humans and certificates of origin can only go so far in these cases.
Some of the tasks found in Microsoft’s blog specifically targeted potential mitigations too, like using LLMs to bypass security features such as two-factor authentication, which is often used to make sure only the right people can use sensitive online services.
Of course, Microsoft and OpenAI say they will continue to close accounts, track future threats and share what they find with the broader cybersecurity community. But the ease with which all these systems can be manipulated, and the enormous growth in availability and use of chatbots like ChatGPT, suggests it might be difficult to prevent these tools affecting the democratic process across the world.
Alex’s week
What I’m reading
Wolf Hall by Hilary Mantel. After she died in 2022, I resolved to read more of her work.
What I’m watching
I’ve been rewatching The Matrix. Some of it has held up well, and it feels relevant given the recent doom-filled prophecies of AI taking over the world.
What I’m working on
I’m trying to learn what mathematics is really about – not a small task – for an upcoming feature.
Alex Wilkins is a Â鶹´«Ă˝ reporter covering artificial intelligence, physics and space. Artificially intelligent is a column that cuts through the hype and looks at what AI is really capable of and what it means for us. You can follow him @AlexWilkins22
