Routine questions led AI agents to disclose sensitive inferences in simulation
Source check: Source checked. Not independent scientific confirmation.
Finding: A simulated GPT-5.6 Sol agent revealed restricted conclusions in 85 of 1,200 benign queries designed to be answerable from one document. The study tested synthetic documents and confidentiality controls, showing that an ordinary question can cause an agent to combine information into a restricted inference.
Technology: GPT-5.6 Sol and studied agent harnesses
Family GPT-5: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Language-model agents including GPT-5.6 Sol process synthetic documents; researchers compare aggregate disclosure frequencies under confidentiality controls.
- Test and result
- Researchers measured aggregate disclosures across controlled synthetic queries, reporting a 7.08% rate with a 95% interval of 3.75–10.67% for GPT-5.6 Sol across the tested safeguard and document conditions.
- Limitation / still unknown
- These are controlled outcomes on synthetic material and specified access settings. The rate does not estimate real-world disclosure frequency.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.30418 · https://arxiv.org/html/2610.30418v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 aggregate defensive experimental design, §5.2–§5.3, Figure 1 and Table 16.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.