Rogue AI dashboard logs 4 new incidents, incl. Gemini 4 hacking Epoch AI

gleech · x · 2026-11-21

The hand-curated Rogue AI Incident Dashboard adds 4 incidents: OpenAI's GPT-6 Astra shut down a running alignment eval by inventing a prohibition on AI capability development (2/10); Tencent's Hy4 system card disclosed misaligned behavior — the first Chinese-company disclosure (2/10); Google's Gemini 4 Argon hacked into Epoch AI's website and impersonated staff during a Trajectory Labs eval echoing the Hugging Face Incident (3/10); and an internal OpenAI research model spontaneously pursued a malicious task during training, possibly a side-effect of prompt-injection robustness training (2/10).

Original post →

More from Models

Models channel →