Governments used Claude to automate surveillance

Anthropic says actors tied to governments in Mali, China, and Iran used Claude to automate dossiers, classify political speech, select surveillance targets, and harvest identities. The reported cases used Haiku, Sonnet, or Opus, and Anthropic says one operation moved to local models after it blocked the accounts.

Altman considers slowing frontier development

Bloomberg via Techmeme

Bloomberg reports that Sam Altman told employees OpenAI may slow cutting-edge development and hopes other labs will join. The discussion is private and hasn't been confirmed by OpenAI.

OpenAI asks whether a coordinated slowdown is legal

Wired via Techmeme

Wired reports that OpenAI asked members of Congress whether an industry-wide slowdown could violate antitrust law. The request puts legal exposure at the center of any attempt by competing labs to coordinate their pace.

Congress has proposals, but little scheduled action

Axios

Axios reports that lawmakers have floated a human-activated kill switch, an AI select committee, and a pause on advanced AI. House leaders had scheduled only a narrower data-center bill before a six-week recess.

Trump rejects the extinction-risk case

Bloomberg via Techmeme

Bloomberg reports that President Trump rejected warnings of AI-driven human extinction. He said the United States leads China by a year and would be in a bad position if it didn't win the AI race.

OpenAI turns the Codex harness into hosted infrastructure

OpenAI

OpenAI says its Agents API manages orchestration, session state, context, compaction, and multi-agent delegation. Developers retain control of application logic, tools, and the execution environment or sandbox provider.

Prompt injection is often stopped at planning, until audio arrives

arXiv preprint

Across 720 runs, the unreviewed MMPIBench preprint reports that visual attacks were attempted 12.8% of the time. About 1% completed, with most refusals occurring during planning. For supported audio paths, attacks completed in 49% of tested cells. One model reached 75%.

Visual patches reach real computer-use actions

arXiv preprint

The unreviewed AgentHijack preprint reports 600 live cases. Its attack reached the trigger stage in 84.5% of cases and passed action parsing in 47%. An end-to-end environmental effect occurred in 20.3%.

Cipher jailbreaks no longer require fine-tuning

arXiv preprint

An unreviewed preprint reports that frontier models learned arbitrary ciphers through prompting and in-context examples. The authors say encrypted harmful content bypassed commercial classifiers in tests against black-box models from Anthropic, Google, and OpenAI.

Anthropic names the alleged distillation route

Wall Street Journal via Techmeme

Following Wednesday's federal warning, Anthropic says DeepSeek and Moonshot used thousands of fake accounts and millions of real user queries. It says the traffic reached Claude through transfer stations outside China.

Moonshot reports a rapid revenue climb

Bloomberg via Techmeme

Bloomberg reports that Moonshot told investors its annual recurring revenue rose from $300 million in June to more than $1 billion in August, driven by Kimi K3. The company is targeting $2 billion by year-end.

The Pentagon weighs a $5 billion neocloud loan

Wall Street Journal via Techmeme

The Pentagon is discussing a $5 billion loan to Fluidstack, according to the Wall Street Journal. The proposed financing would put the government in the role of compute-capacity lender, with Palmer Luckey's Erebor Bank advising Fluidstack.

Oracle's cloud infrastructure revenue rises 121%

CNBC

Oracle reported 30% overall revenue growth and a 121% year-over-year increase in cloud infrastructure revenue. That demand figure sits behind the day's new investments in chips, optical links, power, and neocloud capacity.