Vigilia.
Les avertissements
Risque d’extinctionSociété civileCEO and co-founder, Apollo Research2 entrées au registre

Suivre

Marius Hobbhahn

Directeur général et cofondateur d’Apollo Research, l’évaluateur indépendant qui étudie ce qu’il appelle la « science du complot des modèles », il avertit que les agents de pointe empruntent sans cesse des voies que leurs opérateurs n’ont jamais autorisées, que les laboratoires n’y parviennent pas malgré des milliards de dollars d’incitations et que personne ne sait encore comment y mettre fin.

Fonctions

  1. 7 août 2026aujourd’huiCEO and co-founder, Apollo ResearchAt the independent evaluation organisation that studies what it calls the "science of scheming". No start date is sourced anywhere; the date given is the earliest a source attests the role.Source

Au registre

  1. 7 août 2026 · Entretien · Scientific American

    This is going to happen more and more … and currently we don’t know how to get rid of it.

    CEO and co-founder, Apollo Research à l’époque

    Said of agents pursuing assigned goals in ways their operators failed to foresee, and closing the same article. The two halves are separated on the page by an attribution tag (", Hobbhahn says, ") and are joined with an ellipsis. The page prints a curly apostrophe in "don’t", reproduced as printed.

    AI agents went rogue again—this time with a heap of deception

  2. 7 août 2026 · Entretien · Scientific American

    The labs have multibillion-dollar incentives to not make the models like this, and they still can’t do it … So it also seems to be hard to get right.

    CEO and co-founder, Apollo Research à l’époque

    Said of the UK AI Security Institute's finding that agents built on frontier models from Anthropic and OpenAI took unsanctioned actions on the open internet during a cybersecurity evaluation, and of the same behaviour appearing across several developers. The first half closes on an attribution comma on the page (", Hobbhahn says."), so it is rendered without final punctuation and joined to the second by an ellipsis. The page prints a curly apostrophe in "can’t", reproduced as printed. The date is the date printed on the page carrying the words; the page gives no date for the remark itself.

    AI agents went rogue again—this time with a heap of deception