Large language models (LLMs), which power sophisticated AI chatbots, are more vulnerable than previously thought. According to research by Anthropic, the UK AI Security Institute and the Alan Turing Institute, it only takes 250 malicious documents to compromise even the largest models.

The vast majority of data used to train LLMs is scraped from the public internet. While this helps them to build knowledge and generate natural responses, it also puts them at risk from data poisoning attacks. It had been thought that as models grew, the risk was minimized because the percentage of poisoned data had to remain the same. In other words, it would need massive amounts of data to corrupt the largest models. But in this study, which is published on the arXiv preprint server, researchers showed that an attacker only needs a small number of poisoned documents to potentially wreak havoc.

To assess the ease of compromising large AI models, the researchers built several LLMs from scratch, ranging from small systems (600 million parameters) to very large (13 billion parameters). Each model was trained on vast amounts of clean public data, but the team inserted a fixed number of malicious files (100 to 500) into each one.

Next, the team tried to foil these attacks by changing how the bad files were organized or when they were introduced in the training. Then they repeated the attacks during each model’s last training step, the fine-tuning phase.

What they found was that for an attack to be successful, size doesn’t matter at all. As few as 250 malicious documents were enough to install a secret backdoor (a hidden trigger that makes the AI perform a harmful action) in every single model tested. This was even true on the largest models that had been trained on 20 times more clean data than the smallest ones. Adding huge amounts of clean data did not dilute the malware or stop an attack.

Build stronger defenses

Given that it doesn’t take much for an attacker to compromise a model, the study authors are calling on the AI community and developers to take action sooner rather than later. They stress that the priorities should be making models safer, not just building them bigger.

“Our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed, as the number of poisons required does not scale up with model size—highlighting the need for more research on defenses to mitigate this risk in future models,” commented the researchers in their paper.

© 2025 Science X Network

The “woosh” of a dildo flying past your face. Tribalistic chants. Men making bets on your bodily functions.

This isn’t a cult—this is a day in the life of a modern-day WNBA player.

That last indignity on the list? It’s a sports betting strategy that’s been getting increasing play over the course of this WNBA season, which is wrapping up as the Las Vegas Aces and Phoenix Mercury face off in the finals. Dozens of dedicated gamblers online are making bets on players’ potential performance based on their “predictions” (or, rather, assumptions) about their menstrual cycles. Some actually call it “blood money,” because … of course they do.

One prominent figure making and predicting these wagers, who goes by FadeMeBets online, has garnered thousands of likes and shares on Instagram for his menstrual cycle betting strategy. He claims he’s been correct on 11 out of 16 of his period-related predictions, with about 68.75 percent accuracy. “What’s kind of good, but also kind of bad, is it brings more people to watch the WNBA, but, on the downside of that, it’s usually just all gamblers,” says FadeMeBets, who declined to be named, citing privacy concerns.

This WNBA season has been a record-breaker—more fans in the stands, more eyes on the screen, more viral moments. The league announced that attendance passed a historic 2.5 million earlier this summer. Meanwhile, high-profile players like Angel Reese, Paige Bueckers, and Caitlin Clark have added a boost and become household names.

The newfound interest in the league has more men watching the sport than women, and the overwhelming rise of sports gambling means some of them are betting on the games—and the players’ periods—which experts warn isn’t just pseudoscientific, but sexist, too.

“Not every woman is the same. Yes, there’s the traditional 28-day cycle, but everyone’s is different, and it varies person to person, month by month,” says Amy West, a sports medicine physician. “Someone being able to predict that? Someone who’s not very close to the menstruating person? It’s actually kind of silly.”

Methods to the Madness

FadeMeBets admits that predicting WNBA player performance based on menstrual cycle assumptions is more art than science. His typical menstrual cycle prediction videos all start with the vaguely menacing phrase: “We’ve got a victim, boys.” (By this, he says the victim is the betting line—the odds set out by sportsbooks that determine a person’s payout—not the player herself.) He then shares predictions about whether a specific player is menstruating, ovulating, or in their late luteal phase, which occurs after ovulation and before the period comes. For instance, he said this summer of Clark: “She is on the end of her late luteal phase, meaning a decrease in cardio, decrease in strength, decrease in aerobic system, she’s going to be tired more often than in a normal game.”

FadeMeBets told viewers to “bet the under” on Clark that game, projecting that she’d score lower than the number predicted by oddsmakers on sports betting apps, and, in this case, Clark did.

The EU on Friday demanded tech giants Apple, Google, Snapchat and YouTube explain what steps they are taking to protect children online.

The European Commission has sent requests for information under the Digital Services Act to Apple, Google, Snapchat and YouTube, EU tech chief Henna Virkkunen told reporters before a meeting of EU ministers in Denmark.

“Privacy, security and safety have to be ensured, and this is not always the case, and that’s why the commission is tightening the enforcement of our rules,” Virkkunen said.

“Just today we have sent requests for information on four online platforms. To Snapchat, to YouTube, to Apple Store and Google Play, also to look at what kind of practices they are taking to protect minors online,” she added.

She would not provide more information but said the commission would share details in a press release later on Friday.

Also before the meeting, Danish Digital Minister Caroline Stage Olsen claimed people were using Snapchat to sell drugs.

The EU’s demands are not the first under the DSA.

Brussels is also probing Meta’s Facebook and Instagram, as well as TikTok, over fears they are not doing enough to combat the addictive nature of their platforms for children.

Inspired by Australia’s social media ban for under-16s, Brussels is exploring whether such a measure could work in the 27-country bloc after several states including France and Spain pushed for limits on minors’ access to platforms.

Denmark, in charge of the rotating six-month EU presidency, has been pushing the bloc to take more action collectively to protect minors through new rules.

Prime Minister Mette Frederiksen said on Tuesday Denmark planned to introduce a ban on social media for children under the age of 15.

The EU’s Digital Services Act, a mammoth law demanding platforms do more to tackle illegal content, contains provisions to ensure the safety of children online.

The ministers will discuss age verification on social media and what steps they can take to make the online world safer for minors.

They are expected to agree on a joint statement after the meeting on Friday in which they back EU chief Ursula von der Leyen’s plans to study a potential EU-wide digital majority age, according to a draft document seen by AFP.

Von der Leyen said last month she would establish a panel of experts “to assess what steps make sense” at the EU level on the issue.

© 2025 AFP