Component: Full links from the week beginning on July 14 2025
All the stories we shared in our daily LinkedIn news carousel during the week beginning on July 14 2025.
Is this Elon Musk's strangest investment yet? Find out about the billionaire's latest weird decision as well as cannibal robots, leftist language models, a European cyber-blitzkrieg and humanity's possible last chance to save ourselves from rogue AI in the latest edition of Component, our new daily news roundup.
We'll be sharing the headlines as a carousel on LinkedIn, as well as here on Machine.
To get full links for the week's editions, just visit this page (www.machine.news/component), which we'll start updating every Monday from here on.
Perhaps we'll share the full week's headlines as a newsletter at some point in the future - but let's see how we get on!
Follow Machine on LinkedIn to read daily editions of Component
Today's headlines
Leftie language models: AI bias exposed?
Academics from New York University have set out four reasons why LLMs exhibit a “strong and consistent prioritisation of liberal-leaning values, particularly care and fairness”:
- Liberal-leaning training corpora.
- Reinforcement learning from human feedback (RLHF).
- Liberal dominance in academia.
- Safety-driven fine-tuning practices.
Read the paper: “Amazing, They All Lean Left” – Analyzing the Political Temperaments of Current LLMs.
How to achieve AGI? Say NO to LLMs
Yann LeCun, Chief scientist of Meta AI, has said that achieving AGI requires researchers to abandon generative and probabilistic models, as well as contrastive methods, reinforcement learning and one other very visible form of AI.
“If you are interested in human-level AGI, don’t work on LLMs,” he said.
Hannibal Electric: Rise of the cannibal robots
“Robots can grow bigger, faster, and more capable by consuming materials from their environment and other robots.”
The words of engineers who set a foundation for the creation of robots which grow and self-repair by eating other machines.
They said “machine metabolic processes" will be an “essential part of any sustained future robot ecology.”
Read a paper on the cannibal robots.
Microsoft’s “TinyTroupe” of simulated humans
Microsoft has unveiled a new toolkit which auto-generates realistic “imaginary humans”, gives them memories, opinions, political views, even jobs, then lets them interact autonomously in simulated environments like offices or debates.
TinyTroupe can be used to simulate populations of humans using LLM-powered personas for market research, debates or synthetic data generation.
Humanity’s last chance to save itself from rogue AI?
Most of the major AI firms have put their names to a new paper which argues that monitoring models’ chain of thought is a “unique” opportunity to control models’ behaviour, offering a “new and fragile opportunity for AI safety”.
However, the research warns: “There is no guarantee that the current degree of visibility will persist.”
Which is as scary as it sounds...
Elon Musk’s $980k annual investment in waifus
Elon Musk’s xAI has announced that is hiring two highly paid engineers to work on building Waifus - a Japanese word for an attractive female anime character.
With an annual pay packet of up to $440k, the successful candidates will presumably work on Grok’s controversial “Companions” feature, which lets users interact with... attractive female anime characters. A male character called Chad is expected soon.
Apply on the xAI website here or here.
Slow code: AI tools make developers less efficient, study claims
Researchers at METR (Model Evaluation and Threat Research) have released a study that found developers using AI tools take 19% longer to complete tasks than their unassisted colleagues.
“This significant slowdown goes against developer beliefs and expert forecasts,” they wrote, warning of a “gap between perception and reality” which caused sluggish engineers to believe AI had sped them up by 20%.
Get the story and read the research.
US Department of Defense reshores drone production
The DoD has announced plans to ramp up production of US-made drones using off-the-shelf components to “maintain battlefield superiority”.
“Our adversaries collectively produce millions of cheap drones each year," warned Defense Secretary Pete Hegseth as he announced plans to arm combat units with low-cost drones.
Europe worst hit by global spike in cyberattacks
Check Point has released figures which reveal a sharp 21% rise in global cyberattacks during Q2 2025.
Europe experienced the highest increase, with a 22% year-over-year increase, which researchers said is a “concerning sign of intensifying threat activity in the region.”
Headlines from Wednesday July 16:
A nasty surprise inside Google’s AI email summaries?
A new vulnerability could allegedly let threat actors hide malicious instructions inside an email when a recipient uses the Google Gemini “summarise this email” feature.
The malicious code is white-on-white so the victims can’t see it, researchers warned.
Tim Ward, CEO of ThinkCyber, said: "Even trusted AI tools can be manipulated by prompts hidden in everyday emails.”
Tim Ward, CEO at Redflags from ThinkCyber, said: "This new vulnerability in Google Gemini for Workspace highlights a growing concern; that even trusted AI tools can be manipulated by hidden prompts embedded in everyday emails. As AI becomes more integrated into our working lives, it’s vital that users don’t blindly trust alerts or instructions from AI assistants.
"Always verify any security warnings through official channels and never act on urgent requests without confirming their legitimacy. Attackers are getting smarter and are constantly looking for new ways to bypass traditional phishing defences. Unfortunately, we believe this trend is only going to accelerate over time, making security awareness more important than ever."
Here's a link to the disclosure.
A new Vulnerability Research Institute for the UK
The UK's National Cyber Security Centre (NCSC) has launched a new Vulnerability Research Initiative (VRI) which will hunt bugs in partnership with industry experts.
Sylvain Cortes, VP Strategy at Hackuity told us the launch is "a smart and strategic move" which means Britain will be "able to better identify and understand vulnerabilities affecting software and hardware to strengthen defences."
Teachers say no to GenAI
Educators from across the world have published an open letter calling on schools to avoid using "corrosive" GenAI in the classroom - citing "unacceptable legal, ethical and environmental harms".
"GenAI is a threat to student learning and wellbeing," they wrote. There is insufficient evidence for student use of GenAI to support genuine learning gains."
Will OpenAI’s top-secret web browser be called Aura?
A much-anticipated web browser from OpenAI is codenamed Aura, according to an engineer who found references to the name inside the code of the ChatGPT web app.
The browser will compete with Chrome and will give OpenAI to a treasure trove of user data on the scale that has allowed Google to earn vast sums of money.
News of the browser's imminent release was first revealed by Reuters.
Salt Typhoon hacks the US National Guard
A China-linked threat actor compromised the US National Guard and spent almost a year lurking inside its systems, the Department of Defence has confirmed.
Salt Typhoon is infamous for carrying out the worst telecom hack in US history, compromising routers and switches at major firms gain persistent access to American networks, according to government officials.
“Suffering toasters”: A new way to assess AI sentience?
A philosopher has claimed the best way to work out whether an AI is conscious is to isolate it in sensory deprivation (SD) and see how it reacts.
“If a machine displays signs of distress... we can attribute it to an internal process of the mind,” Ira Wolfson of The Hebrew University of Jerusalem writes.
I suffer, therefore I am...
Read the full paper: Suffering Toasters - A New Self-Awareness Test for AI
Medieval medicine was “weirdly similar to TikTok trends”
We have mouth taping and frog venom. They had lizard shampoo and peach mashing.
British Academy researchers have released new details of weird medieval medicines and claimed dark age health hacks are comparable to TikTok wellness trends.
"People were engaging with medicine on a much broader scale than had previously been thought," said Meg Leja, an associate professor of history at Binghamton University who specializes in the political and cultural history of late antique and medieval Europe. "They were concerned about cures, they wanted to observe the natural world and jot down bits of information wherever they could in this period known as the ‘Dark Ages.’"
"A lot of things that you see in these manuscripts are actually being promoted online currently as alternative medicine, but they have been around for thousands of years."
Filefix social engineering attack tested in the wild
Check Point Research has issued a warning about a social engineering attack, dubbed "FileFix".
The new threat loads a PowerShell command into the user’s clipboard which executes when pasted into the Explorer address bar.
Eli Smadja of Check Point said: "This technique doesn’t rely on complex exploits, but on manipulating routine user behaviour."
Read Check Point's blog about FileFix.