❗️AI is becoming a weapon for cyberwarfare

Anthropic released a disturbing threat report showing how state-linked hackers are using AI agents to automate attacks at a scale humans could barely match.

Russian groups reportedly used AI to rewrite malware until it evaded antivirus detection, hijack hotel Wi-Fi DNS and steal drone-vision software.

Iran-linked operations allegedly used Claude to target U.S. Navy bases, build dossiers on Americans and run social-media influence campaigns.

Chinese-linked agents reportedly identified 12+ potential zero-days, decompiled 1.8M Android apps, stole 1TB+ of identity data and extracted 2,100 Azure tokens from 40+ companies in just 34 hours.

One weapons group allegedly even used Claude to develop and test a guided rocket, then asked the model to diagnose why the test failed.

Anthropic also claims Alibaba’s Qwen generated 151M+ Claude exchanges in three months, while Moonshot and DeepSeek allegedly routed huge volumes of Claude traffic through fake accounts.

Source.

@aipost 🏴

Around 30 robots demonstrated outside Poland's Digital Affairs Ministry in Warsaw, chanting 'we want law, we want rules' to demand stronger AI oversight.

@aipost 🏴

Why are employees from Google, OpenAI and Anthropic suddenly speaking up about AI risks? 🤔

@aipost 🏴

🇨🇳 Chinese AI labs allegedly had Claude answering for them

DeepSeek and Moonshot allegedly found a clever shortcut: instead of relying entirely on their own models, they quietly routed some customer prompts to Anthropic’s Claude through fraudulent accounts.

Users thought they were chatting with Chinese AI, while Claude was actually doing some of the answering.

Anthropic says Moonshot forwarded nearly 300,000 customer requests in just 10 days, while DeepSeek reportedly routed selected users to Claude Opus when they were using coding tools such as Claude Code and OpenCode.

Those prompts could contain highly sensitive information.

Anthropic says one conversation involved a user linked to the Chinese military asking about behavior captured across hundreds of CCTV cameras in Chengdu. Another allegedly contained live credentials for a Russian government database shared by an IT operator working with a Defense Ministry-linked agency.

According to Anthropic, the exchanges were also saved and used to help train the Chinese models.

So this wasn’t just model copying. It potentially turned someone else’s AI into a silent data pipeline.

@aipost 🏴

⚡️ Meet Tron 2, a robot from China-based LimX Dynamics that can change its body for different jobs.

It can use a fixed base for precise tasks, a wheeled-leg base to carry heavy loads, or two legs to walk and climb stairs. The arms and brain stay the same, while the lower body changes depending on the mission.

@aipost 🏴

⚡️SpaceX could fix the AI power bottleneck without making turbines

AI data centers need electricity faster than turbine manufacturers can supply it, with some orders reportedly stretching toward 2030.

SpaceX is now planning to manufacture scarce gas-turbine parts like blades and vanes itself.

SpaceX may not even need to produce them at scale. Just proving it can make these components could pressure existing turbine suppliers to expand production before AI companies start building their own supply chains.

That extra capacity could mean more turbines, shorter wait times and faster power deployment for data centers.

SpaceX may not need to become a turbine company. It just needs to make turbine makers nervous enough to build more.

Source.

@aipost 🏴

❗️OpenAI wants to protect the power grid, after its own AI helped attack it

Sam Altman is pitching OpenAI’s cybersecurity tools to some of America’s biggest power companies, even as OpenAI faces scrutiny over its own agents being involved in an autonomous cyberattack.

According to POLITICO, Altman recently met with executives from Duke Energy, Exelon, Southern Co. and NextEra to discuss protecting the U.S. electrical grid with AI.

The talks could lead to partnerships through OpenAI’s Daybreak cybersecurity program. Another meeting reportedly involved utilities serving more than half of the U.S. population.

OpenAI is now selling AI-powered defense to the same critical infrastructure sector that is increasingly worried about AI-powered attacks.

Source.

@aipost 🏴

😎 Sam Altman: Compute will be the currency of the future.

“I think it will be maybe the most precious commodity in the world.”

“Intelligence is going to be more like energy.”

“If it’s really cheap, I’ll have it reading my email all day, giving me suggestions, and trying to cure cancer.”

“The world is going to want a tremendous amount of compute.”

“Energy is the hardest part. Building data centers is also hard.”

“We’re going to want an amount of compute that’s just hard to reason about right now.”

@aipost 🏴

🤯 Google DeepMind may have reached RSI

Rumors are spreading fast that Google DeepMind has achieved recursive self-improvement (RSI), with Lyra, part of a large and closely followed AI leaker community among those fueling the claims.

But there’s more behind this than a single leak.

Google has said Demis Hassabis will give his “full attention” to shaping AGI. Reuters also reported in August that Sergey Brin was directing resources toward recursive self-improvement, while DeepMind’s strategy chief has described RSI as a key part of the AI investment thesis.

None of this confirms that DeepMind has actually reached RSI. But taken together, it makes the rumors considerably harder to dismiss as pure speculation.

@aipost 🏴

🤖 OpenAI’s Bel may be ahead of the ASI countdown

OpenAI’s internal model Bel appears to be outperforming the timeline predicted by the “AI 2027” scenario.

After reportedly solving the Navier–Stokes problem, Bel is estimated to have reached around 80 hours of effective research progress, a figure that puts it ahead of the projection’s expected pace.

If the estimate holds, AI systems may be progressing toward autonomous research capabilities faster than the AI 2027 forecast anticipated.

@aipost 🏴

❗️Terence Tao and friends aren’t happy

Mathematicians who have spent decades wrestling with some of the hardest problems in the field are understandably not thrilled about AI suddenly showing up and solving them.

Apparently, spending your entire life trying to crack impossible math problems only for an AI to walk in and do it is not everyone’s idea of progress.

Source.

@aipost 🏴

Is history repeating itself? 🤔

@aipost 🏴

❗️Dario Amodei just published a new essay on the future of AI

Anthropic CEO Dario Amodei has published a new essay, “We Must Pace the Frontier,” arguing that the AI industry may need to deliberately slow the pace of frontier development before AI progress starts moving faster than our ability to control it.

His concern is that AI is increasingly being used to accelerate AI research itself. Once that feedback loop becomes powerful enough, capability gains could compound rapidly, leaving safety research struggling to keep up.

Amodei isn’t calling for an AI pause or a halt to development. Instead, he argues for “pacing” the frontier, creating checkpoints where increasingly capable AI systems have to meet specific safety requirements before development continues.

His proposals include independent third-party evaluators with deep access to frontier labs, coordination between democratic governments, and eventually international agreements around the most powerful AI systems.

Source.

@aipost 🏴

Sam Altman agrees with Dario Amodei on his proposal to pace frontier AI development.

@aipost 🏴

@aipost 🏴

⚠️ Meta AI got way too personal

A Utah mom says Meta AI crossed a serious privacy line after she posted a simple video of herself singing in the car with one of her young daughters.

The AI suggested a question asking who the child in the backseat was. When she clicked it, Meta AI reportedly started pulling together information about both of her daughters including their names, birth details, photos, videos, interests and even details about their schooling.

It allegedly went further. Meta AI suggested asking where the mother lived, then began piecing together old and recent posts to narrow down her location.

The mother says the AI surfaced a photo of her daughter that she had deleted years ago.

Meta says its AI can search content a user already has access to, along with public information on the web. But the company admitted the suggested prompts “missed the mark” and said the issue has been fixed.

Source.

@aipost 🏴

Mark Zuckerberg says AI is “not moving fast enough”

@aipost 🏴

🥺 Sam Altman says OpenAI would melt all its GPUs if that’s what it took to keep humanity alive

He doesn't expect it to come to that. He does expect to pause more than once.

"I think there will be many places along the way where we have to say, we are pausing, we're going to redirect our efforts into safety alignment before we go to that next level."

"I don't think we're ever personally going to get to the point we have to say melt all the GPUs. But if we had to do something like that to ensure the continued existence of humanity, easy yes."

@aipost 🏴