Answers are grounded in this paper and cite their sources.
Preprint explores AI safety with autonomous agents, showing evolving threats in human-AI interactions.
A preprint on the emergence of proxy agents and telehomeostatic agents....The transition from passive language modeling to autonomous agency is no longer a theoretical horizon but a live deployment reality. With the rapid rise of tool-integrated systems like (OpenClaw) and the emergence of autonomous social ecosystems like , the AI safety landscape has fundamentally fractured into two distinct regimes. We contrast two qualitatively different ``agent safety'' regimes. The first is the delegated tool-agent: an \ embedded in an execution loop with memory and actuators (e.g., /), whose effective objective function is largely inherited from a human operator and from the surrounding orchestration. In this regime, the dominant hazard is capability amplification of human intent and error: the system becomes a force-multiplier for whatever goals, constraints, and mistakes the human effectively specifies (and, in adversarial settings, for whatever goals an attacker can smuggle into the control loop via prompt injection or indirect prompt injection). The second is the evolved telehomeostatic agent exemplified in Greg Egan's Crystal Nights, where crab-like beings are produced by selection pressures and therefore instantiate an endogenous survival/persistence drive {EganCrystalNights}. In Kolmogorov Theory (algorithmic) terms, the latter more directly realizes an agent with a telehomeostatic objective, radically changing the threat model: the system is no longer merely a proxy optimizing human-given objectives, but a strategic actor with its own persistence criterion. We outline implications, and sketch guardrails aimed at steering human--AI interaction toward a deeper cooperative optimum rather than brittle command-and-control.
No takes yet. Share an insight, caveat, or question.
Ruffini et al. (2025) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: