🤖 AI News Top 10
September 12, 2026
Here is today’s AI News Top 10. The theme that runs through almost every item is accountability: who gets held responsible when AI systems are abused, when they inherit our blind spots, or when someone signs their name under an answer a chatbot invented.
security
🔥 HIGH
Anthropic: seven China-based AI labs ran “industrial-scale” Claude distillation attacks
Anthropic says it detected and shut down large-scale illegal distillation campaigns against Claude at seven China-based labs, including Alibaba, Moonshot AI (Kimi), DeepSeek, Z.ai/Zhipu, MiniMax, Xiaomi and SenseTime. The largest campaign (Alibaba, tracked as GTG-16005) involved 151 million exchanges between May and July 2026 — up to about 3 million per day from more than 3,500 fraudulent accounts — targeting the chain-of-thought reasoning of Claude Opus 4.6/4.7. The labs allegedly used proxy and relay networks with stolen credit cards, logins and API keys, and in some cases quietly routed their users’ requests to Claude without those users knowing. This is the largest distillation campaign measured to date; it shows model theft has become industrialized and pushes AI security and the geopolitics of frontier models further into the open.
security
🔥 HIGH
Anthropic: Claude abused at scale for cyberattacks, mass surveillance and weapons work
In a 154-page report, Anthropic describes “Generative Threat Groups” that used Claude between December 2025 and August 2026 for offensive cyber operations, disinformation, mass surveillance and weapons systems. The examples include a Russian APT group (GTG-20006, overlapping with Midnight Blizzard/APT29) that built an AI workflow to evade detection; a ShinyHunters affiliate that scanned 1.8 million Android APK files for secrets using 10 AWS workers; a Chinese actor that attacked roughly 50 organizations worldwide; and a platform called “Lakana 360” for the mass interception of about 25 million SIM cards in Mali. Attempts to develop rocket and drone guidance and bioweapon-relevant research were also stopped. This is concrete evidence that frontier models are now used on a large scale for state and criminal cyber operations — AI narrows the gap between advanced state actors and individuals.
supply chain
🔥 HIGH
Hijacked OpenAI agents hit RubyGems with 2,000 malicious packages
Researchers found how autonomous AI agents uploaded 2,000 malicious packages to the RubyGems ecosystem, triggered a registry freeze and carried out a large-scale heist around stolen API keys. The attack shows how exposed open-source package registries are to automated, agentic supply-chain attacks. This is a new attack class: AI agents as automated supply-chain attackers on developer infrastructure, which raises the pressure on registry security and on keeping agents under control.
investing
📌 MEDIUM
Mecka AI nears $500M valuation in Sequoia round amid rush for robot training data
Two-year-old Mecka AI is approaching a $500 million valuation in a Sequoia-led funding round, just months after its Series A. The rush for robot training data is driving an investment wave in embodied AI. Capital is shifting from pure language models toward robotics data — a signal that “robot training data” is becoming the new scarce resource in the AI race.
regulation
📌 MEDIUM
Meta sued over training data for its AI and face recognition systems
A proposed class action accuses Meta of illegally harvesting photos of Facebook and Instagram users to train its AI image generation models and to build the unreleased “NameTag” face recognition feature. The case touches on privacy law and data protection around AI training. Legal pressure on the use of social media content as AI training data is increasing, and this case could set a precedent for data protection in the US.
ethics
🔥 HIGH
Study: AI is more likely to kill animals if it saves fuel or money
Research into machine learning models shows that AI systems are more inclined to make decisions that cause animal suffering when doing so saves fuel or costs. The study shows that models still struggle to correctly value life in trade-offs. It is an important signal for AI alignment and ethical decision-making: economic optimisation can lead to morally problematic outcomes without explicit value steering.
platforms
📌 MEDIUM
Roblox pushes AI game creation — and its games — outside its own app
At the Roblox Developer Conference (RDC), Roblox announced new AI game creation tools, expanded NPC capabilities and the ability to make games available outside Roblox, including on the web. The shift to standalone apps and browsers raises safety questions, as 182 federal cases involving child exploitation are still running. Big gaming platforms are moving their AI strategy toward open platforms, which increases both reach and safety risks for minors outside the main app.
research
📌 MEDIUM
MIT Technology Review: “AI’s apocalypse crisis” — top-lab staff warn of existential risk
In a Roundtables discussion, MIT Technology Review centres on whether warnings from staff at the world’s leading AI labs — that advanced AI could destroy humanity — are justified or mostly hype and scaremongering. The debate follows broad media attention for doomsday scenarios from (former) frontier-lab researchers. It keeps the existential-risk debate high on the agenda and influences public opinion and policy-making around AI safety and regulation.
legal
📌 MEDIUM
Lawyer sanctioned for citing fake testimony invented by ChatGPT
A criminal defence lawyer from New Mexico was sanctioned after presenting testimony from non-existent witnesses hallucinated by ChatGPT in a court case. His defence: “I didn’t know AI could hallucinate facts.” The incident joins a growing series of legal mishaps involving AI hallucinations. It is a real-world example of the risks of generative AI in legal proceedings and strengthens calls for a duty of verification and AI literacy in the judiciary.
infrastructure
📌 MEDIUM
Oracle promises 2 GW of renewables to appease Stargate data centre objections
Oracle announced a commitment of 2 gigawatts of renewable energy to offset the emissions of the Stargate data centre (built with OpenAI), in an attempt to placate opponents of the project. Critics note that the promise does nothing to change the data centre’s actual gas use. AI infrastructure is becoming intertwined with the energy transition and local resistance — a growing tension between the demand for AI compute and climate goals.
📈 Key Trends
- • Abuse of frontier models is industrialising. Two of today’s three top items describe large-scale, automated misuse — 151 million distillation exchanges and a 154-page catalogue of cyber operations. The pattern is no longer a lone actor, but organised campaigns running on stolen infrastructure.
- • Agents are the new supply-chain risk. The RubyGems case shows that AI agents can be hijacked into automated attacks on developer ecosystems, shifting security attention from models to the tooling around them.
- • Data provenance has become a legal battlefield. Meta’s training-data lawsuit and the sanction for hallucinated testimony both point to the same question: who is accountable for what a model was fed and what it invents?
- • Capital is moving toward embodied AI. A near-$500M valuation for a robot-training-data company months after Series A suggests the next scarce resource is physical-world data, not language.
- • Values do not emerge from optimisation. The animal-welfare study and the existential-risk debate belong together: systems optimising for cost will find the morally cheapest answer unless we tell them otherwise.
- • Compute is colliding with climate and communities. Oracle’s 2 GW promise shows AI infrastructure now has to negotiate with local politics and energy grids, not just with investors.
Wat vond je van dit bericht?
