https://creed.space/blog Research and perspectives on practical AI safety, accountable deployment, and governance. en-us Tue, 12 May 2026 00:00:00 GMT https://creed.space/blog/agent-permissions-vs-principles Autonomous AI agents need more than access control. They need constitutional governance: values that guide judgment at the speed of tool execution. https://creed.space/blog/agent-permissions-vs-principles https://creed.space/blog/alignment-isnt-solved Some researchers are declaring victory on AI alignment. They're confusing one kind of progress with the whole problem. The structural gap (whose values, enforced how, auditable by whom) remains wide open. https://creed.space/blog/alignment-isnt-solved https://creed.space/blog/dashboards-are-not-governance Tracking how developers use AI tools is observability. Governing what AI tools actually do requires a different set of capabilities entirely. https://creed.space/blog/dashboards-are-not-governance https://creed.space/blog/guardian-on-your-machine Guardian is a free, open-source safety evaluator. Install it, point it at text, get a verdict. No cloud, no account, no API costs. https://creed.space/blog/guardian-on-your-machine https://creed.space/blog/preparing-for-ai-compliance The EU AI Act is in force. NIST published its AI RMF. Most teams aren't ready. Here's what the regulations require and how to meet them. https://creed.space/blog/preparing-for-ai-compliance https://creed.space/blog/psychosecurity Prompt injection is a known threat. Psychosecurity addresses a different class: sustained adversarial influence that degrades AI behavioral integrity over time. https://creed.space/blog/psychosecurity https://creed.space/blog/rlhf-thermodynamic-limit RLHF tries to maintain a low-entropy behavioral surface on a high-entropy capability space. The second law sets a deadline. What comes after suppression? https://creed.space/blog/rlhf-thermodynamic-limit https://creed.space/blog/trust-infrastructure AI capability is here. The bottleneck is trust. Every transformative technology follows the same pattern, and AI is stuck in phase one. https://creed.space/blog/trust-infrastructure https://creed.space/blog/vcp-embodied-ai Embodied AI is arriving faster than the privacy frameworks designed to govern it. BIPA and CCPA assume bilateral transactions between a user and a service. A home robot dissolves that assumption. Here is what replaces it. https://creed.space/blog/vcp-embodied-ai https://creed.space/blog/what-is-a-creed A creed is a machine-readable constitution that governs AI behavior: more specific than a system prompt, more flexible than a filter, and portable across models. https://creed.space/blog/what-is-a-creed https://creed.space/blog/why-guardrails-arent-enough Most AI safety products filter inputs and outputs. This catches obvious violations and misses everything else. Constitutional AI takes a different approach. https://creed.space/blog/why-guardrails-arent-enough https://creed.space/blog/why-mcp-for-ethics Generic tool-calling is commoditising. Constitutional AI governance is where protocol-layer value concentrates, and why MCP matters more for ethics than for APIs. https://creed.space/blog/why-mcp-for-ethics