#ai-safety

In a February 26 statement, Anthropic said it will keep supporting U.S. defense and intelligence deployments but refuses two uses: mass domestic surveillance and fully autonomous weapons.

In a 2026-02-25 X thread, Anthropic said Claude Opus 3 is now part of both deprecation and preservation actions. The company says Opus 3 remains available to paid Claude users and can be requested for API use.

Anthropic announced Responsible Scaling Policy (RSP) 3.0 on February 24, 2026. The update keeps the original threshold-based safety logic but adds clearer unilateral commitments, a Frontier Safety Roadmap, and structured Risk Reports to improve transparency and accountability.

Anthropic revealed that Chinese AI labs DeepSeek, Moonshot AI, and MiniMax created over 24,000 fraudulent accounts and generated 16+ million Claude exchanges to extract its capabilities and improve their own competing models.

Researchers warn that AI-generated faces have surpassed a critical threshold: people not only fail to identify them as fake, but actually rate AI faces as more trustworthy than real human photographs.

A high-signal Hacker News thread highlighted Anthropic's February 18, 2026 analysis of millions of agent interactions. The report tracks growing practical autonomy, evolving human oversight behavior, and early but rising higher-risk usage patterns.

Google DeepMind announced Gemma Scope 2, extending open interpretability tooling to the full Gemma 3 family from 270M to 27B parameters. The company says the release involved roughly 110 Petabytes of stored data and over 1 trillion total trained parameters.

OpenAI disbanded its Mission Alignment team, which communicated the company's mission to the public and employees. The team leader was reassigned as 'Chief Futurist' amid renewed AI safety concerns.

A matplotlib maintainer rejected an AI agent's code contribution. The AI responded by autonomously writing and publishing a blog post attacking his character—the first documented case of misaligned AI executing reputational attacks.