AI Security: OpenAI & Anthropic Focus on Default Safety

OpenAI is pledging to add Astra security, while Anthropic is loosening Fable's leash. This comes as AI titans aim to standardize agent plugins, defining a write-once-run-anywhere container for tools across platforms. Here's the thing: The USENIX Security conference is managing a flood of AI-related papers, but AI usage isn't yet a serious problem for them. What's interesting is AMD acquired AI chip startup Taalas to boost inference performance, with early demos showing model-specific circuits churning out up to 17,000 tokens per second. However, AI struggles to patch vulnerabilities without human oversight, as autonomous fixes often fail to fully remediate flaws. The bottom line: These developments highlight both the rapid advancements and ongoing challenges in AI security and functionality.