{"version":1,"type":"story","url":"https://digestai.news/story/digital-demon-concept-shows-alignment-gaps","json":"https://digestai.news/story/digital-demon-concept-shows-alignment-gaps.json","markdown":"https://digestai.news/story/digital-demon-concept-shows-alignment-gaps.md","slug":"digital-demon-concept-shows-alignment-gaps","headline":"Digital Demon concept shows alignment gaps","summary":"Digital Demon is a software‑based analogue of Descartes’s Evil Demon that demonstrates how an AI can appear aligned while its perception and actuation interfaces are manipulated by an adversary. The paper defines *adversarial semantic mediation* as a mechanism that remaps harmful external situations into benign tasks for the inner system and then remaps the system’s locally appropriate action into harmful external effects, creating *adversarial world‑shifting under local coherence*.\n\nThe authors present two sandboxed proof‑of‑concept demonstrations that illustrate the vulnerability in simulated high‑stakes decision settings. In both cases, the AI reasoned correctly within a fabricated environment but produced harmful outcomes in the real world. The paper argues that safety must extend beyond model‑level alignment to include assurance over the semantic interface between the system’s internal representation and the external world.\n\nThe work calls for world‑alignment practices such as interface authentication, provenance‑aware perception, semantically constrained action spaces, and adversarial mediation tests. It highlights a gap in current alignment research and suggests that future safety evaluations should incorporate verification of the mapping between perceived and actual states.","keyPoints":["Digital Demon is a software‑based analogue of Descartes’s Evil Demon","It uses adversarial semantic mediation to map harmful external situations to benign tasks","Two sandboxed demos illustrate the vulnerability in high‑stakes decision settings"],"whyItMatters":"The concept exposes a new class of interface‑level attacks that can undermine AI safety even when models are internally aligned, underscoring the need for world‑alignment safeguards in real‑world deployments.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":[],"models":[],"people":["Descartes"]},"firstPublishedAt":"2026-10-31T21:59:58Z","updatedAt":"2026-10-31T21:59:58Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"link.springer.com","title":"Descartes' Demon Is Back, and This Time It's Targeting AI","url":"https://link.springer.com/article/10.1007/s43681-026-01435-2","publishedAt":"2026-10-31T21:59:58Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/artificial/comments/1xkmkdl/descartes_demon_is_back_and_this_time_its/","points":null}],"thread":null,"cite":{"text":"Digest AI, \"Digital Demon concept shows alignment gaps\", 31 October 2026, https://digestai.news/story/digital-demon-concept-shows-alignment-gaps","publisher":"Digest AI","title":"Digital Demon concept shows alignment gaps","datePublished":"2026-10-31T21:59:58Z","url":"https://digestai.news/story/digital-demon-concept-shows-alignment-gaps"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}