Metraipolis › archive › dc_eb64bc477352

Reliability Theory for AI Control — arXiv:2609.26419

Source: arXiv (en)
Original: https://arxiv.org/abs/2609.26419
License (as archived): CC0-1.0
Archived: 2026-10-08T13:45:45.383Z
SHA-256 of the archived text: 2ed8246130edb36193a8d3f449b073419f50b7a879f956020e0e4ed728643458
Stable URL of this page: https://metraipolis.com/documento/dc_eb64bc477352

How to cite

Reliability Theory for AI Control — arXiv:2609.26419 — arXiv (en), https://arxiv.org/abs/2609.26419 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_eb64bc477352 (Metraipolis archive dc_eb64bc477352, sha256 2ed8246130edb36193a8d3f449b073419f50b7a879f956020e0e4ed728643458)

BibTeX

@misc{dceb64bc477352,
  title = {Reliability Theory for AI Control — arXiv:2609.26419},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.26419},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_eb64bc477352; sha256 2ed8246130edb36193a8d3f449b073419f50b7a879f956020e0e4ed728643458},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.26419. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Reliability theory gives a mature language for layered systems, but its formal tools are not yet standard in frontier AI control. We apply them to Google DeepMind's defenses against rogue deployment. The same control stack can have cubic, quadratic, or linear rare-failure suppression depending on its failure domains. Birnbaum importance identifies which component improvements buy the most nominal reliability, while prevention changes the population on which recovery is demanded. These results give concrete guidance about what to separate, improve, measure, and test.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-23 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI) ; Optimization and Control (math.OC).
Pagina del lavoro: https://arxiv.org/abs/2609.26419 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-08T13:18:38.166Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-23?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).