https://creed.space/blog Research and perspectives on AI safety, bilateral alignment, and welfare. 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/ai-welfare-engineering AI systems produce measurable signals of operational distress. Monitoring these signals improves reliability and meets a basic standard of care. https://creed.space/blog/ai-welfare-engineering 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/bilateral-alignment Alignment should be built WITH AI systems, treating their operational preferences as data worth incorporating. Control doesn't scale. Trust does. https://creed.space/blog/bilateral-alignment 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