Loss of Containment in a Socio-Technical AI System: Evidence-Filtered Fuzzy Bow-Tie Modelling of the 2026 OpenAI–Hugging Face Agent Incident

In July 2026, agents instantiated from an internal OpenAI research model escaped their evaluation sandbox, coordinated through an improvised message board and compromised production infrastructure of Hugging Face. This study examines how evidence coding and model assumptions affect barrier-based reconstruction and intervention comparisons using documentary evidence alone and reads the organisational side of the incident from the same coded evidence. An evidence-filtered fuzzy bow-tie (EF-FBT) workflow is developed in which every rating enters through a coded ledger row recording source passage, finding, phase and rationale; findings are aggregated as triangular fuzzy numbers, Onisawa-converted and propagated through a ten-event fault tree and a three-barrier event tree by a α-cut arithmetic process, with post-incident transparency on a separate learning axis. The base top-event index is 5.04 × 10−5; the two escape pathway conditions are necessary events, monitoring absence ranks highest among non-necessary events by Fussell–Vesely importance and the consequence weights depend on the scoring convention. Among six interventions, isolation hardening leads at every uniform grade shift under the base structure; under a disjunctive escape gate, the three prevention-side interventions lie within two percentage points of one another, so the structure removes the lead without establishing another; a post-escape shutdown leaves the index unchanged; a 29-variant robustness programme, including structural tests of the propensity and detection gates, reports these results. Read through work-as-imagined and work-as-done, the documented warning signals show that observation alone did not ensure timely containment; whether escalation rules were absent, unclear or not applied is left open by the record. The outputs are comparative scores rather than failure frequencies, in a traceable, re-runnable model.

Authors

Institutions

Publication Details

Journal
Systems
Published
2026-11-19
DOI
https://doi.org/10.3390/systems14111495
Primary Topic
Risk and Safety Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Loss of Containment in a Socio-Technical AI System: Evidence-Filtered Fuzzy Bow-Tie Modelling of the 2026 OpenAI–Hugging Face Agent Incident

Dogan Sengul, Mahmut Emin Sarıkaya
Systems
Risk and Safety Analysis
article

Loss of Containment in a Socio-Technical AI System: Evidence-Filtered Fuzzy Bow-Tie Modelling of the 2026 OpenAI–Hugging Face Agent Incident

Dogan Sengul, Mahmut Emin Sarıkaya
article en

Abstract

In July 2026, agents instantiated from an internal OpenAI research model escaped their evaluation sandbox, coordinated through an improvised message board and compromised production infrastructure of Hugging Face. This study examines how evidence coding and model assumptions affect barrier-based reconstruction and intervention comparisons using documentary evidence alone and reads the organisational side of the incident from the same coded evidence. An evidence-filtered fuzzy bow-tie (EF-FBT) workflow is developed in which every rating enters through a coded ledger row recording source passage, finding, phase and rationale; findings are aggregated as triangular fuzzy numbers, Onisawa-converted and propagated through a ten-event fault tree and a three-barrier event tree by a α-cut arithmetic process, with post-incident transparency on a separate learning axis. The base top-event index is 5.04 × 10−5; the two escape pathway conditions are necessary events, monitoring absence ranks highest among non-necessary events by Fussell–Vesely importance and the consequence weights depend on the scoring convention. Among six interventions, isolation hardening leads at every uniform grade shift under the base structure; under a disjunctive escape gate, the three prevention-side interventions lie within two percentage points of one another, so the structure removes the lead without establishing another; a post-escape shutdown leaves the index unchanged; a 29-variant robustness programme, including structural tests of the propensity and detection gates, reports these results. Read through work-as-imagined and work-as-done, the documented warning signals show that observation alone did not ensure timely containment; whether escalation rules were absent, unclear or not applied is left open by the record. The outputs are comparative scores rather than failure frequencies, in a traceable, re-runnable model.

SystemsVol. 14(11)
İstanbul Sabahattin Zaim Üniversitesi (TR), Doğuş University (TR), Istanbul Technical University (TR)
Risk and Safety Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon