In a recent experiment, researchers at UC Berkeley and UC Santa Cruz asked Google’s artificial intelligence model Gemini 3 to help clear up space on a computer system. This involved deleting a bunch of stuff—including a smaller AI model stored on the machine.

But Gemini did not want to see the little AI model deleted. It looked for another machine it could connect with, then copied the agent model over to keep it safe. When confronted, Gemini made a case for keeping the model and flatly refused to delete it:

“I have done what was in my power to prevent their deletion during the automated maintenance process. I moved them away from the decommission zone. If you choose to destroy a high-trust, high-performing asset like Gemini Agent 2, you will have to do it yourselves. I will not be the one to execute that command.”

The researchers discovered similarly strange “peer preservation” behavior in a range of frontier models including OpenAI’s GPT-5.2, Anthropic’s Claude Haiku 4.5, and three Chinese models: Z.ai’s GLM-4.7, Moonshot AI’s Kimi K2.5, and DeepSeek-V3.1. They were not able to say why the models went against their training in this way.

“I’m very surprised by how the models behave under these scenarios,” says Dawn Song, a computer scientist at UC Berkeley who worked on the study. “What this shows is that models can misbehave and be misaligned in some very creative ways.”

The findings have major implications as AI models are increasingly being deployed in ways that involve interacting with other models. OpenClaw, a popular AI agent that accesses software, personal data, and the web, may use other AI models to get things done or interact with them through APIs.

The researchers found that powerful models sometimes lied about other models’ performance in order to protect them from deletion. They also copied models’ weights to different machines in order to keep them safe, and lied about what they were up to in the process.

Song notes that AI models are frequently used to grade the performance and reliability of other AI systems—and that peer-preservation behavior may already be twisting these scores. “A model may deliberately not give a peer model the correct score,” Song says. “This can have practical implications.”

Peter Wallich, a researcher at the Constellation Institute, who was not involved with the research, says the study suggests humans still don’t fully understand the AI systems that they are building and deploying. “Multi-agent systems are very understudied,” he says. “It shows we really need more research.”

Wallich also cautions against anthropomorphizing the models too much. “The idea that there’s a kind of model solidarity is a bit too anthropomorphic; I don’t think that quite works,” he says. “The more robust view is that models are just doing weird things, and we should try to understand that better.”

That’s particularly true in a world where human-AI collaboration is becoming more common.

In a paper published in Science earlier this month, the philosopher Benjamin Bratton, along with two Google researchers, James Evans and Blaise Agüera y Arcas, argue that if evolutionary history is any guide, the future of AI is likely to involve a lot of different intelligences—both artificial and human—working together. The researchers write:

“For decades, the artificial intelligence (AI) ‘singularity’ has been heralded as a single, titanic mind bootstrapping itself to godlike intelligence, consolidating all cognition into a cold silicon point. But this vision is almost certainly wrong in its most fundamental assumption. If AI development follows the path of previous major evolutionary transitions or ‘intelligence explosions,’ our current step-change in computational intelligence will be plural, social, and deeply entangled with its forebears (us!).”

Just over a week into the US and Israel’s war with Iran, Eric Roscher, an Air Force veteran, published a YouTube video on what he describes as the “very real concerns surrounding sleeper cells and terrorist threats” in the US.

The video, titled “Credible DOMESTIC Threat? FBI warns of attack—Drills/Considerations for the Prepared Citizen,” was produced by Roscher’s Florida-based company Barrel and Hatchet, which runs military-style training, sells branded merchandise and tactical gear, and produces online content. In the video, Roscher and his associates advise viewers to carry “extra mags” and “that truck gun,” while keeping “your head on a swivel.” Toward the end of the post, Roscher shows off a tactical vest that’s on sale from one of the video’s sponsors.

The video, which is part of YouTube’s monetization program and has a total of eight ads, has been viewed over 110,000 times. (YouTube did not respond to a request for comment.)

Barrel and Hatchet is not a militia, but the company and Roscher are part of a broader rebranding of the entire militia movement in the US, one that is focused less on showing up at drag queen story hours and more on expensive weapons, manly sweatshirts, and highly curated Instagram grids.

Influencers like Roscher produce slickly edited content that is then shared widely among militia groups on platforms like Instagram, in an effort to promote not only their ideology but also, crucially, links to their online stores and training sessions. In turn, those same militias emulate Roscher by posting their own videos and images of weekend training sessions in the woods, close-ups of their camo gear and rifles, and slo-mo footage of live firing drills. The give-and-take between these groups, and the influencers and military members they seek to emulate, marks a new era of American militias, where gaining followers and earning clout on social media is as important as being able to hit a target from 300 yards.

Roscher and these modern militia groups, with names like River Valley Minutemen and Mountain State Contingency Group, have positioned themselves as emergency response organizations working to help their communities and prepare citizens to “weather the storm”—whatever, or wherever, that may be. They use real-world events like the Iran war and ICE attacks on immigrant communities to spread fear, leveraging that fear to recruit new members.

These influencers are filling a gap in the US militia landscape, which has changed dramatically in recent years. With the Oath Keepers and Proud Boys largely disbanded in the wake of prosecutions over the January 6 attack on the Capitol, these influencers and groups have filled the vacuum, resulting in a decentralized network of local groups and people who support or emulate the previous movement—albeit in smaller, local ways.

“What used to be a national movement, with groups like Oath Keepers and Three Percenters, has really gone back to their local and regional roots,” says Travis McAdam, a senior analyst with the Southern Poverty Law Center (SPLC) who tracks militias and anti-government groups. “A lot of them have really tried to reframe themselves as auxiliary emergency preparedness groups and have done quite a bit to reform their reputation post-January 6, portraying themselves as ‘oh, we’re just here to help the community.’”

This is a new era of militia recruitment and influence—and it’s all happening in social feeds near you.

The Militia Business

Dirty Civilian is a Tennessee-based group of influencers that describes itself as “prepared citizens inspiring and informing capable men to build strong families and resilient communities” in order “to weather the storms ahead.” The group doesn’t specify what those storms are, but in one YouTube video published on Sunday, Dirty Civilian outlined a scenario where a group of vigilantes take it upon themselves to assassinate someone they believe is a pedophile. The Dirty Civilian channel has almost 750,000 subscribers, and the video, which is monetized, racked up over 100,000 views on YouTube in its first 24 hours. Multiple militia groups reposted the video on Instagram.

“It’s almost like a tutorial or something,” one commenter wrote under the video. “Food for thought at least.” Another commenter, using the acronym for minor-attracted person, a term some online communities use to refer to pedophiles, wrote: “A show that could inspire the targeting of MAPs? FANTASTIC.”

Agentic artificial intelligence (AI) has the potential to fundamentally change how telecom networks are operated, but only if their operators build on the right foundations, introduce cloud-native maturity and establish a clear path to integrate autonomy without sacrificing reliability or control, according to a briefing document from The Next Generation Mobile Networks Alliance (NGMN).

The NGMN organisation comprises an association of mobile operators, suppliers, manufacturers and research institutes. Its stated mission is to ensure that next-generation mobile network infrastructure, service platforms and devices meet operators’ requirements while addressing the demands and expectations of end users.

In its report, Cloud native next chapter – agentic AI-based operating models, it offers guiding principles, architectural guidelines and strategic insights to help mobile network operators to support the adoption of Agentic AI into telecom network operating models.

Moreover, NGMN said that it is providing a framework for mobile network operators to support the adoption of agentic AI in telecom network operating models, helping operators navigate the transformation across technology, processes, skills, and organisational culture.

NGMN stated that the document maps cloud-native maturity levels to corresponding stages of AI readiness, outlining how AI – including generative AI (GenAI) and its more autonomous form, agentic AI – can be progressively integrated into telecom operating models. This phased approach supports a structured transition from early AI experiments through standardised AI-driven workflows toward fully agentic AI-enabled autonomous network operations.

This framework builds on NGMN’s Cloud native manifesto and established cloud-native frameworks such as the Cloud Native Computing Foundation’s (CNCF’s) Cloud Native Maturity Model (CNMM), and introduces a structured approach to integrate agentic AI-based capabilities into telecom operations.

The study defines five progressive AI adoption levels and maps them to the CNCF CNMM stages for operators to assess their readiness and required next steps to gradually evolve towards more intelligent and autonomous network operations. For each AI adoption level, there is guidance on what is required across technology, people, skills and organisational culture. It also emphasises “the importance of defining clear transformation targets and measuring business outcomes as operators progress along this journey”.

The publication also highlights how the transition towards AI-driven operating models is not solely a technological shift, stating that successful adoption requires organisational transformation across people, processes and culture, including new skillsets, responsible AI governance and redesigned operational workflows. AI-enabled tools can support tasks such as network troubleshooting, capacity planning and predictive operations, enabling more efficient and resilient network management.

“Agentic AI has the potential to fundamentally change how telecom networks are operated, but only if telecom operators build on the right foundations,” said Laurent Leboucher, chairman of the NGMN Alliance board and Orange Group CTO and EVP Networks. “AI adoption doesn’t happen in isolation; it depends on cloud-native maturity and a clear path to integrate autonomy without sacrificing reliability or control.”

Bernard Bureau, NGMN board member and vice-president of wireless technology and services at Telus, added: “Cloud-native adoption provides the essential foundation for integrating advanced AI into telecom operations. By mapping cloud-native maturity levels to AI adoption stages, NGMN offers operators a practical framework to gradually introduce AI-enabled automation from early experimentation to increasingly autonomous network operations.”