Skip to main content

A Fundamental Flaw Leaves LLMs Strikingly Vulnerable To Attack

21 hours 7 minutes ago
joshuark quotes a report from MIT Technology Review: It is impossible to make large language models fully secure against hacks because of a fundamental flaw in how they work, a team of researchers argue in a paper presented at the International Conference on Machine Learning, a top AI conference, this month. The claim has huge implications for the safety of this technology. By taking advantage of this flaw, which concerns how LLMs identify who or what is giving them instructions, the researchers were able to make popular LLMs spit out information they had been trained not to provide, such as how to synthesize cocaine and how to sabotage a commercial aircraft's navigation system. "There's a real probability that this is going to be a problem that's fundamentally unsolvable," says Charles Ye, an independent researcher and coauthor of the ICML paper. [...] The ICML paper describes attacks against several of OpenAI's models, but Cui and Ye say that they have since seen similar results with models made by Anthropic, Alibaba, and DeepSeek. Cui and her colleagues wanted to find out why an attack like chain-of-thought forgery was so effective. They suspected it had something to do with the mechanism that LLMs use to keep track of where their instructions are coming from. But what Cui and her colleagues discovered is that LLMs are in fact very bad at keeping track of different roles. In a series of experiments that looked at what was going on inside a handful of different models, the researchers found that LLMs seem to identify the role of a specific chunk of text not by the tags around it but by the style of that text and the words it contains. The upshot, the researchers claim, is that all an attacker needs to do to hack an LLM is write text that spoofs a certain role. And because roles are a fundamental part of how LLMs work, no amount of training will fully solve the problem. "There's going to be a huge economic incentive for people to do jailbreaks and prompt injections," says Cui. The best defense could be to expect the worst. Organizations shouldn't trust LLMs, and they should expect that anything done by agents could be unsafe, he says: "That's not a great solution, but it just might be what we have to do." "It's really incredible that these things are being deployed everywhere to control super-critical systems. There's been no study of the fundamental science here. We're all doing it ad hoc."

Read more of this story at Slashdot.

BeauHD

Microsoft's $450 Billion Jump Is Biggest In Stock Market History

22 hours 7 minutes ago
Microsoft shares surged as much as 17% after reporting 43% growth in Azure revenue, putting the company on track to add a record $490 billion in market value in a single day. Bloomberg notes that it "would eclipse Nvidia's $440 billion addition, following President Donald Trump's announcement of a 90-day tariff pause last year, as the biggest ever." From the report: The nearly $500 billion jump is larger than the market capitalization of roughly 96% of S&P 500 stocks, data compiled by Bloomberg show. It's also bigger than the combined value of the benchmark's 44 smallest members, which includes companies like Domino's Pizza Inc., Clorox Co. and Hasbro Inc. Microsoft's one-day add in value also dwarfs many of the world's other equity markets. South Africa, Turkey, Finland and Vietnam all have total stock market values that are less than what the software maker is set to add on Thursday.

Read more of this story at Slashdot.

BeauHD

ABC Accuses FCC of 'Attempted Censorship'

23 hours 7 minutes ago
ABC accused FCC Chair Brendan Carr of "attempted censorship," arguing that an early review of its eight broadcast licenses is politically motivated retaliation over the network's coverage and could chill the entire media industry. "The retaliation against ABC is a signal to every media company in the country: accommodate the Administration's view of what news coverage should look like or pay the price," the Disney-owned television network wrote. Politico reports: "Across the government, regulatory and contracting carrots and sticks have been trained on other disfavored speakers," ABC's lawyers added in the 119-page filing. "The tools vary; the objective does not: a media industry too fearful of official reprisal to report the news freely." Carr has repeatedly rejected accusations of censorship while defending his efforts to rein in what he calls abusive, politically motivated behavior by licensed TV broadcasters. "I don't view the FCC as the speech police," he told POLITICO this week for an upcoming episode of the podcast "The Conversation." In its filing, ABC pointed to an outpouring of support it has received in more than 152,000 comments in the agency's license review, as well as recent warnings from conservative Supreme Court Justice Neil Gorsuch, Sen. Ted Cruz (R-Texas) and various pro-free-market organizations about the dangers of a politically motivated FCC. Those include letters from a bipartisan group of former FCC leaders, community and advocacy groups and lawmakers of both parties. "The Commission's attempted censorship of ABC has attracted condemnation across the political aisle," the network wrote.

Read more of this story at Slashdot.

BeauHD