---
title: "#AI Safety - AFeng Blog"
url: "https://afengblog.com/en/tag/ai-safety"
language: "en"
total: 2
---

# Tags: AI Safety

- [Black October 2026 — How PHP2Shell and LiteLLM Exposed the AI Open-Source Crisis](https://afengblog.com/en/blog/black-october-2026-php2shell-litellm-ai-open-source.html) (2026-11-03 · Security & Privacy): Between October 17 and 30, 2026, 14-year-old PHP2Shell RCEs, Google Brotli out-of-bounds writes, TianoCore UEFI flaws, eMudhra's single-byte CA key, and the 101-minute LiteLLM supply-chain attack detonated concurrently. Here is how frontier AI models compressed decade-old bug hunting into a single week—and pushed a 13-developer open-source foundation to the brink.
- [Inside GPT-6 Rosalind — Same Weights as Astra, While OpenAI Charges 50% More to Relax Safety Guardrails](https://afengblog.com/en/blog/gpt-6-rosalind-same-weights-selling-permissions.html) (2026-11-03 · AI Exploration): OpenAI prices its flagship life sciences model GPT-6 Rosalind Research at $15/$75 per million tokens—a 50% premium over general-purpose GPT-6 Astra—yet its official system card admits both models share 100% identical weights and zero new training data. By tracing how cutting biosecurity refusals from 29.2% to 9.4% unlocks latent capability, this analysis unpacks how frontier AI labs turn safety sandboxes into zero-marginal-cost licensing valves and outlines a hybrid routing architecture that trims API spend by 33.3%.