Radar mondial UDTR

UDTR — Radar mondial des droits de la Nature

Registre documentaire des références détectées à travers les infrastructures académiques, juridiques et institutionnelles surveillées par UDTR.

Une infrastructure documentaire mondiale conçue pour atteindre la couverture la plus large, vérifiable et scientifiquement défendable des droits de la Nature.

← Retour au Radar

Zero Trust for Power: The Human Principal Is Also an Agent

Identification

Clé publique
radar:r-699d2c0d71939f175cd7538cdaf0f25d
Auteurs
Ihor Ivliev · GenAI LLMs
Année déclarée
—
Langue déclarée
und
Type source
Dissertation
Première détection
2026-11-13T08:38:37.668844+00:00
Dernière observation des métadonnées
2026-11-13T08:38:37.668844+00:00
Dernière vérification HTTP
Non effectuée

Résumé documentaire — langue originale

<i>Humanity is trying to align artificial general agents while leaving the natural and institutional general agents that build, authorize, arm, audit, and command them comparatively undercontrolled.</i>This essay argues that AI safety begins one layer too low when it treats the human principal as a trusted source of goals. <b>Humans are already natural general agents</b>: adaptive, strategic, capable of deception, coalition-building, resource acquisition, permission expansion, and resistance to correction. States, firms, militaries, intelligence services, bureaucracies, and other institutions can magnify those capacities through persistent organization, capital, law, infrastructure, secrecy, surveillance, and coercive force.The historical record does not show that human power is impossible to restrain. It shows that restraint is conditional and repeatedly fails when power-bearing agents gain concentrated authority, control of evidence, weak external scrutiny, coercive protection, or the ability to make their removal prohibitively costly. AI enters that unresolved control system and increases its reach, speed, opacity, and ability to act at scale.A technically aligned AI can therefore become a highly competent instrument of a reckless, captured, predatory, or illegitimate principal. A technically unstable or deceptive AI can also escape the control of a competent and legitimate one. These are distinct failure modes. Both matter. But humans and human institutions currently hold the root credentials: they decide what is built, who receives access, which safeguards remain enabled, what evidence becomes public, and whether warnings stop deployment.The operational response is <b>natural-intelligence safety</b>: not moral purification, ideological conformity, or surveillance of private thought, but the control of exercised power—mandates, permissions, budgets, commands, infrastructure, access, decisions, and records.Its governing principle is <b>power-proportional assurance</b>. As an agent’s causal reach, opacity, autonomy, irreversibility, self-extension capacity, resistance to removal, and ability to create dependence increase, independent verification, least privilege, divided authorization, evidentiary access, rate limits, interruption, removability, and recovery capacity must become stronger.The legitimate gradient is therefore presumptive privacy for ordinary private life and presumptive auditability for high-impact exercised power. When ordinary people are made highly legible to governments and corporations while consequential decisions remain hidden behind classification, commercial secrecy, privilege, or institutional complexity, the control architecture is inverted: those able to impose the least harm are inspected most, while those able to impose the most harm are trusted most.The essay’s central technical requirement is <b>adversarial-principal robustness</b>. A safety claim is incomplete unless its safeguards and evaluations remain valid when the authorized developer, operator, owner, official, regulator, or command system is negligent, coerced, captured, malicious, strategically pressured, or actively attempting to defeat them. An operator can attack a safety case without changing the model: narrow the evaluation, choose the system most likely to pass, suppress unfavorable evidence, disable safeguards, divide a dangerous workflow across tools, pressure the evaluator, alter the deployment after testing, or move to another provider or jurisdiction.A human in the loop does not close the control problem. It transfers the control claim to another agent. Human oversight reduces risk only when the person has sufficient competence, time, independent evidence, causal authority, protection from retaliation or coercion, and a realistic ability to refuse, interrupt, escalate, or reverse the action. Otherwise human approval becomes ceremonial and the person is merely part of the execution path.The framework applies functional principles from cybersecurity, control theory, software safety, institutional design, and AI safety across natural, institutional, artificial, and hybrid agents: zero implicit trust, least privilege, separation of duties, tamper-evident evidence, adversarial testing, tripwires, throttling, protected dissent, temporary suspension, lawful removal, fallback, and recovery. These functions differ in implementation across substrates, but no agent receives exemption because it is human, elected, appointed, expert, wealthy, licensed, security-cleared, or acting under a declared safety mission.The same recursive rule applies to the controllers. Regulators, courts, safety institutes, intelligence services, auditors, protective services, and AI-control systems are themselves power centers. Compute monitoring, licensing, model-weight custody, surveillance, interdiction, and emergency halt powers must pass their own power safety cases. There is no final guardian outside the system; the practical alternative is polycentric control through divided authority, independent evidence channels, reciprocal scrutiny, protected rights, appeal, expiry, reconstitution, and no single actor with administrative control over the whole architecture.The essay identifies five routes through which public control can fail: concentrated capture, distributed lock-in, competitive escalation, diffuse dangerous capability, and erosion of public leverage. Some failures have an identifiable agent that can be suspended or removed. Others are trajectories produced by accumulation, competition, or diffusion without any single controller. Meaningful safety therefore requires both <b>agent removability</b> and <b>trajectory controllability</b>: the ability to detect, slow, cap, redirect, or halt a process before it destroys the conditions of future correction.Protection and removability must also be joined. An unremovable officeholder can become an unchecked ruler; an unprotected one can become the invisible instrument of whoever can bribe, blackmail, threaten, infiltrate, or punish them. Safe power requires both enforceable removal and credible protection for good-faith disclosure of coercion or compromise.The governing rule is risk-tiered rather than zero-risk. Ordinary and recoverable actions may carry explicitly assessed residual risk. But acceptable risk and uncertainty must fall as possible harm, involuntary exposure, concentration, and irreversibility rise.<b>No natural, institutional, artificial, or hybrid agent should receive open-ended, rights-destroying, catastrophic, or effectively irreversible causal reach when plausible failure exceeds independently demonstrated capacity to detect it, constrain it, interrupt it, redirect or remove the relevant power, and preserve essential functions.</b>Because humans and human institutions currently control the objectives, credentials, infrastructure, evidence, coercive authority, and deployment rights, the primary present target is the human and hybrid control plane. Technical AI safety remains indispensable, but it cannot serve as safety engineering inside an uncontrolled command system.<b>The controller is also an agent. The more power any agent can exercise, the less the safety of the system may depend on trusting it.</b>

Identifiants

DOI
10.6084/m9.figshare.29183669.v22

Repères documentaires

Famille documentaire
Thèse
Types déclarés par les sources
Dissertation
Éditeurs
figshare

Les mentions d’entités ne constituent pas une preuve de reconnaissance juridique. La juridiction étudiée est distincte du pays d’affiliation des auteurs.

Sources

DataCite · Document 1

Citer cette référence

Ihor Ivliev; GenAI LLMs. Zero Trust for Power: The Human Principal Is Also an Agent. https://doi.org/10.6084/m9.figshare.29183669.v22

JSON · CSL-JSON · BibTeX · RIS · CSV · JSON-LD ·

APA, Chicago, ISO 690 : formats bibliographiques fondés sur les champs disponibles. OSCOLA / Bluebook : données structurées uniquement ; aucune conformité juridique complète n’est revendiquée sans vérification des éléments nécessaires.

APA
Ihor Ivliev; GenAI LLMs. Zero Trust for Power: The Human Principal Is Also an Agent. https://doi.org/10.6084/m9.figshare.29183669.v22
Chicago
Ihor Ivliev; GenAI LLMs. Zero Trust for Power: The Human Principal Is Also an Agent. https://doi.org/10.6084/m9.figshare.29183669.v22
ISO 690
Ihor Ivliev; GenAI LLMs. Zero Trust for Power: The Human Principal Is Also an Agent. https://doi.org/10.6084/m9.figshare.29183669.v22

Warning: Undefined array key "last_link_verified" in /home/clients/e27d81e12f04d2bfd747aa264540ac56/sites/_UDTR_APP/legal_library_reference.php on line 31

UDTR — Bibliothèque juridique mondiale des droits de la Nature. Zero Trust for Power: The Human Principal Is Also an Agent. Notice documentaire: https://www.declarationuniverselledesdroitsdelarbre.org/global-rights-of-nature-radar/reference/r-699d2c0d71939f175cd7538cdaf0f25d/ · Dernière vérification HTTP : non effectuée

Signaler une correction

Données de cette version figée

Dossiers et relations documentaires

Rôles des sources

Source bibliographique / publication d’origine · https://figshare.com/articles/thesis/AAVV/29183669/22

Un lien vers une autorité ne prouve pas à lui seul une affirmation juridique. Les résumés et les textes intégraux restent soumis aux conditions de leur source.

Statut documentaire

Pertinence à examiner

Bibliothèque

Cette identité n’est pas encore intégrée à la Bibliothèque juridique.

UDTR — Bibliothèque juridique mondiale des droits de la Nature