AI Safety and Governance
Safety in this corpus ranges from model-level interventions to deployment governance and contested moral claims about agents. The sources agree more strongly on the need for measured safeguards, scope boundaries, and evaluation than on the ontology or long-term policy conclusions behind them.
Synthesis
- The Assistant Axis work reports a model-internal safety intervention, while Claude Code auto mode describes operational filtering and permission gating; both are first-party evaluations with stated limitations. [src] [src]
- The open-weights position argues for capability-sensitive testing and targeted controls rather than a blanket ban, making it a policy stance rather than settled evidence. [src]
- The Yara shutdown highlights high-risk health settings where wellness, clinical care, crisis response, and longitudinal observation can blur; its proposed guardrails do not remove those governance questions. [src]
- The model-welfare proposal is explicitly an unverified, normative view. It belongs in the corpus as a debate about system design, not as evidence that models have personhood. [src]