AI news, 29 August: Court rules Pentagon’s Anthropic blacklist unlawful

US judge overturns Anthropic supply-chain blacklist. A federal judge in California ruled on 27 August that the Trump administration’s designation of Anthropic as a national-security supply-chain risk was unlawful, according to reports from Reuters, The Verge and Politico. US District Judge Rita Lin found the Pentagon’s actions under Defense Secretary Pete Hegseth amounted to unconstitutional retaliation against the company for its public stance on military AI uses, violating the First and Fifth Amendments, and called the measures “illegal and baseless.” The ruling vacates the designation and related contractor bans; the government is expected to appeal. The decision is a major win for the Claude maker in its dispute over AI safety “red lines.”

Over 100 firms warn of coming AI cyberattacks. OpenAI, Anthropic, Google, Microsoft and more than 100 other companies, including cybersecurity and financial firms, issued a joint open letter on 27 August calling for urgent collective action against AI-enabled cyber threats, TechCrunch, Axios and the BBC reported. They warned that attacks will become far more widespread and sophisticated within months, putting hospitals, water systems and internet infrastructure at risk, and urged governments and organisations to strengthen defences, share tools and fund protection for critical services while there is still a “limited window.”

Anthropic research shows AI agents fixing alignment issues. Anthropic published findings that automated systems and Claude-based agents developed training methods mitigating ten common alignment failures across all tested benchmarks without harming overall performance, according to coverage from TechCrunch and Unite.AI. The work points to progress in AI helping improve its own safety, though researchers note the tests remain relatively narrow.

Anthropic advances IPO preparations. The company plans to publicly unveil its IPO prospectus after Labor Day, with a potential listing in late September or early October, The Information reported, citing people familiar with the matter. Earlier coverage noted high valuation discussions and a large addressable-market pitch to investors.

Other developments. Hugging Face unveiled Microduck, a roughly $399 open-source 25 cm robot for AI and robotics experimentation. Google DeepMind introduced double-blind benchmarking protocols to improve evaluation trust. Reports also noted continued infrastructure spending, including Lambda securing $1 billion in debt to buy Nvidia chips. Nvidia’s reported near-$13 billion Hugging Face talks remain a closely watched business story but saw no major confirmed closing update in the latest window.