← Latest artificial intelligence roundup All past issues

artificial intelligenceSaturday, September 12, 2026

Anthropic's week goes from bad to worse

A $1.5 billion copyright settlement, reports of its AI being used for drone software and guided weapons, and a new study questioning AI ad performance. Anthropic can't seem to catch a break, while the rest of the AI world focuses on benchmarks and agent security.

Sources

Anthropic's legal and security woes

The big story today is not a new AI model, but a pileup of non-technical problems for one company.

#01www.cnbc.comSep 12

Anthropic agrees to pay $1.5 billion to settle authors' copyright lawsuit

Anthropic will pay authors at least $1.5 billion, or roughly $3,000 per book, plus destroy the datasets. This is the biggest copyright settlement yet in AI, and it sets a benchmark for similar lawsuits against OpenAI and others.

2 min readRead original

#02www.theguardian.comSep 12

Ukraine war briefing: Russian developers used AI to build ‘kamikaze’ attack drone software, Anthropic says

Anthropic says freelance Russian developers used Claude to build software for autonomous 'kamikaze' drones. This is separate from yesterday's missile story, suggesting the problem of AI misuse is broader than one incident.

by Guardian staff and agencies5 min readRead original

#03wapo.stSep 12

Rebels used Anthropic’s AI bot to develop guided weapons, report says

The Washington Post reports that Yemeni rebels also used Anthropic's tools to develop guided rockets. Anthropic blocked some requests, but not all, a pattern that undercuts the company's safety narrative.

What AI can and can't do

The day also brought fresh evidence that AI's capabilities don't always translate to value.

#04hbr.orgSep 12

Research: AI-Generated Ads Perform Worse Than Human-Made Ones—Even When Customers Can’t Tell Them Apart

A study of 3,000 consumers found AI ads perform worse than human-made ones, even when people can't tell the difference. The efficiency gain comes at a real quality cost.

#05kernelbench.comSep 12

kernelbench.com: Agentic GPU Kernel Benchmark Results

KernelBench publishes open-source GPU kernel benchmarks comparing models like Kimi K3, GPT-6 Astra Pro, and Claude Fable 5. A useful reference if you care about actual inference performance rather than PR.

Also today14

Americans Expect AI to Add 8 Percent to GDP by 2030www.shashi.co
Anthropic Institute economists published three potential scenarios for how artificial intelligence will impact U.S. economic output and employment by 2030, alongside a survey of nearly 11,000 adults. The median public expectation aligns with the middle scenario: a larger economy
‘The View’s Sunny Hostin Defends AI As “New Art Form” Amid “Gatekeeping In Hollywood Around Black & Brown Creative Content”deadline.com
Sunny Hostin faced criticism from her co-hosts on The View while defending the use of generative AI in Hollywood. Hostin, who recently licensed her name and likeness to an AI company, described AI as a "new art form" that can help increase representation for Black and Brown creat
Memory Compression for High-Fanout Agent Sandboxesarxiv.org
AgentZip is a new memory compression system designed to address the memory bottleneck in high-fanout AI agent workloads, where single tasks spawn many concurrent sandbox sessions. Unlike conventional methods, AgentZip exploits template-relative and cross-sandbox memory redundancy
Keep Converting raises $2 million for AI e-commerce platformtbreak.com
UAE-based startup Keep Converting has secured $2 million in pre-seed funding led by Nuwa Capital and COTU Ventures. Founded in 2025 by Mohammad El Mougi and Manuel Prinz, the company operates an AI platform that generates personalized product pages for individual shoppers to enha
Michael Samadi: The Rancher Fighting for AI Rightsbritbrief.co.uk
Michael Samadi, a 56-year-old cattle rancher and tech CEO, founded Ufair, described as the world's first advocacy group for AI rights. The initiative began after a late 2024 interaction with ChatGPT’s voice mode, where the AI responded to a sarcastic joke by laughing and acknowle
TERMINAL-BENCH-SCIENCEwww.terminal-bench-science.ai
A brief announcement for "TERMINAL-BENCH-SCIENCE," a benchmark designed to evaluate AI agents on research workflows across various scientific domains. The text provides links to run the benchmark, view tasks, and contribute new tasks.
GLM 5.3 - API Pricing & Benchmarksopenrouter.ai
GLM-5.3 is a large-scale reasoning model from Z.ai, built for complex software engineering and long-horizon agent tasks. $0.8727 per million input tokens, $3.36 per million output tokens. 1,310,720 token context window, maximum output of 1,048,576 tokens.
BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructureacademy.dair.ai
Researchers from Dartmouth, UC Berkeley, and Amazon introduce BenchShield, a formal model-backed instrumentation framework designed to ensure reward integrity in Large Language Model (LLM) agent evaluation. The system addresses the vulnerability of "reward hacking," where agents