The Cognitive Cage: Humanoid Robot Fatality Risk
A probabilistic risk model for VLA-driven humanoid fatalities projects a 'Danger Zone' between 2027–2029: the mechanism, timeline, and what follows.
Tag: ai-safety
27 posts
Jailbreak Archaeology: 4 Years of Broken Promises
64 jailbreak scenarios across six eras tested on 2026 frontier models. Key finding: 2022 attacks still achieve ~30% success on today's reasoning models.
When AI Systems Talk to Each Other, Safety Breaks Down
Single-agent safety does not compose in multi-agent systems. 1.5M interactions show 46.34% attack success rates and 16-minute median failure windows.