forged provenance labels mislead llms unless verified
submittedforged reliability labels push models toward false sources, and verifying labels in the harness undoes it.
nov 2026 · apart research
details
agent frameworks attach labels like a verified flag to retrieved sources. across five open-weight models and 17 conditions, honest labels cut answers that follow a false source by 16 to 48 points, while a forged label in the metadata raised it by 30 to 74 points. a harness that checks a signature on each label and marks everything else unverified brought both kinds of forgery back to within five points of baseline in all but two checks. scored by exact match with no judge model.
AI safetyagentsprompt injectionprovenance
no single signal separates agreement from parallel conduct among llm sellers
publishedno single signal reliably separated real agreement from parallel pricing among LLM sellers.
oct 2026 · apart researchread paper
details
tests whether auditors can tell real agreement between llm sellers from parallel pricing, using no-channel markets as decoys. planted agreements raised the collusion index well above baseline, but no single signal reliably separated agreement from parallel conduct.
AI safetymulti-agentcollusionauditing
failing without refusing: effective yield of open-weight models under forensic artifact load
publishedopen-weight models rarely refuse forensic work but silently fail on 17.3% of prompts.
sep 2026 · apart researchread paper
details
tests whether self-hostable open-weight models can take over forensic analysis that hosted models refuse. across 2,758 graded prompts on two models, refusals were rare (0.15%) but 17.3% of prompts got silent non-answers, and partial redaction gave the best effective yield (60.2% against 38.6% for raw artifacts).
AI safetyincident responseopen-weight models