Anthropic Fable 5 Returns; AI Safety Standards & Threats

Anthropic's Fable 5, an AI model, is back online after being inaccessible for 19 days. Amazon researchers mitigated a jailbreak technique with a new safety classifier, aiming to block such attempts in 99% of cases. The White House is expected to release new voluntary standards for AI frontier models this week. These standards are likely to be treated as requirements by AI companies. Executive Order 14409 comes into force in early August. OpenAI has delayed three versions of GPT-5.6 to meet these voluntary standards, pushing back their release by at least two to three weeks. Google introduced Gemma 4 12B, a lightweight model designed to bring agentic multimodal intelligence to laptops. It supports audio-visual inputs and runs on 16GB of VRAM. A study by Ramp suggests that firms adopting AI can grow their headcount by 10.2% to 12% over two years, challenging the idea that AI will eliminate jobs. This growth primarily applies to high-intensity adopters and engineering-intensive, venture-backed companies. Check Point Research demonstrated that the DeepSeek AI model can generate functional ransomware from a prompt. This ransomware runs entirely within a browser, requiring no exploit or technical expertise. The operator of the RustDuck Botnet is reportedly rewriting it in.