Insights, perspectives, and lessons learned from building AI safety systems at scale.
Humanity is approaching the Great Filter. We face four vectors of existential risk from unchecked AI: permanent power concentration, gradual human disempowerment, catastrophic pandemics enabled by frontier models, and critical infrastructure collapse. The path forward requires curving capability with alignment.
OpenClaw has shipped over 90 security fixes in two weeks — all addressing traditional software vulnerabilities. None touch the structural gap: no boundary between external content and trusted instructions, and no monitoring between the model’s decisions and tool execution. The problem isn’t in the code. It’s in the intelligence layer.
Part 2 of the Astroware Security Research Series. Existing defenses operate before the agent turn begins. Once a skill passes static analysis, there is nothing between that skill’s output and the agent acting on it. We make the case for bidirectional guard architecture — two guards, two positions — and why it belongs in core.
Subscribe to our blog to receive the latest insights on AI safety and security.
Subscribe