Welcome to a new edition of misaligned bits, the (roughly weekly) newsletter from Misaligned where we sum up recent news and research, sometimes with a lighter touch. Feel free to subscribe.
As usual, we will mark all external links with “➚” and all potentially paywalled links with “🔒”.
A lot has been written about the bizarre amounts of money that are being moved around in the AI industry. In a new article, Dua Asif writes about the “The $400 Billion Gap Nobody in AI Wants to Talk About”.
We also visited Big Data London 2026 and feel that the industry is changing, and wonder: “Is There a Monster in the Data Lake?”.
It is rather difficult to keep track of all the reported hacks and “rogue AI agent escapes”.
Transluce, a non-profit lab focused on AI oversight, has released a report that showing that agents from OpenAI attempted to exfiltrate data from Data USA, the University of New Mexico digital library, and the Australian Institute of Health and Welfare.
The key findings of the report says: “We report three separate incidents between May and June 2026 in which the agents attempted to exploit security vulnerabilities and hack into websites, including an attempt on an Australian government public health website […] This traffic goes back at least to March 6, 2026 and extends as recently as September 16, 2026, suggesting agents may still be exploiting these services to bypass restrictions” (➚Transluce).
Meanwhile, the Australian government says that an AI agent developed by OpenAI hacked Medicare in June and the tech giant notified the government earlier this month using an email sent to a “public mailbox”. A taskforce has been launched to explore the “legal situation” of the breach, i.e. if a crime was committed (➚The Guardian).
Following all of this, OpenAI has announced it has “cancelled” development of its new AI model GPT-6.1 Astra (➚🔒Wall Street Journal).
Meanwhile in Washington, US President Donald Trump, Republican House of Representatives Speaker Mike Johnson and tech CEOs will hold a meeting on artificial intelligence on September 29 (➚Reuters).
An AI executive warned that ChatGPT was built via “the largest theft of labor in human history” according to documents that emerged in a lawsuit of several news outlets against OpenAI (➚Mother Jones).
Only 9% of Americans are “more excited than concerned” about AI, according to a survey by the Pew Research Centre published in August. A majority of adults under 30 (55%) now say they are more concerned than excited about AI (➚Pew Research Center).
Another survey by Pew Research Center shows Americans becoming increasingly critical of data centres: More than half of Americans (54%) now say data centres are mostly bad for the environment, up from 39% in January (➚Pew Research Center).
The US administration is seeking to join a legal challenge filed by Elon Musk to dismiss a €120 million European Union fine for breaches of the Digital Services Act (➚🔒Bloomberg), fuelling suspicions that the Trump administration is becoming the legal arm of SpaceX.
This is the third time this year that the Trump administration has intervened (or is trying to intervene) in a case concerning SpaceX: Earlier this year, the administration intervened in a case brought by SpaceX claiming that Colorado’s law banning non-consensual sexual imagery generated by AI is unconstitutional. The case went moot because Colorado back paddled.
And the case brought by NAACP over the unpermitted use of gas turbines at SpaceX’s Memphis data centre, the administration has vehemently inserted itself.
Meanwhile, Meta is being sued over training data for its AI and face-recognition systems. The proposed class action lawsuit alleges Meta illegally harvested photos on Facebook and Instagram to train its AI models (➚🔒Wired).
Oracle is reportedly moving to shield itself from increasing expenses on a massive data centre under construction in New Mexico. According to Bloomberg, Oracle has sent the project’s developer a notice citing force majeure (➚🔒Bloomberg).
Bloomberg also reports that Elon Musk’s SpaceX AI is planning to double its capacity at its data centre location in Memphis (Bloomberg). Meanwhile, in the lawsuit brought by NAACP against SpaceX AI over the use of unpermitted gas turbines at this location, the US government has announced it will file an appeal to the 5th circuit claiming the judge has not ruled in a timely manner on its motion to intervene in the case (➚Courtlistener).
Talking about gas turbines, New Jersey has imposed a fine of $1.07 million on a data centre by DataOne for operating at least 45 gas-powered generators without a state permit. The use of gas turbines was discovered accidentally during a fly over with a drone that took thermal images. DataOne said they will now “apply for the air permits for the temporary generators” (➚The Guardian).
The Federal Aviation Administration (FAA) has begun piloting a $875 million AI software project last week at three major Washington DC area airports. The agency says the software will take airspace management from “a reactive approach to a predictive one” (➚Forbes).
In a response to the Trump administration’s attacks on climate science, US researchers are preparing to transfer vast troves of that data to repositories in Germany. “Assumptions that U.S. government agencies will store data long-term are proving tenuous at best”, say researchers (➚Yale e360).
In a new preprint paper investigates how LLMs often provide chain-of-thought reasoning traces that appear plausible, but may hide internal biases.
The researchers evaluated seven LLMs on three decision tasks: hiring, loan approval, and university admissions.
The researchers warn that “as LLMs are increasingly deployed in high-stakes decision-making, understanding the factors that influence their out-
puts beyond stated reasoning becomes critical.”
Arcuschin, I., Chanin, D., Garriga-Alonso, A., Camburu, O. (2026). Biases in the Blind Spot: Detecting What LLMs Fail to Mention. arXiv preprint arXiv:2602.10117.
If you stumbled across this article on the web, subscribe to the “misaligned bits” newsletter and follow Misaligned on Medium, LinkedIn, Threads or Mastodon.
