In a vignette study of 354 U.S. participants, deletion-based privacy controls outperformed all other controls in increasing willingness to engage with GenAI chatbots for emotional support, while technically complex controls like local-only processing and model training opt-outs reduced engagement.
hub
Can llms keep a secret? testing privacy implications of language models via contextual integrity theory
15 Pith papers cite this work, alongside 12 external citations. Polarity classification is still indexing.
15
Pith papers citing it
12
external citations · Pith
abstract
The interactive use of large language models (LLMs) in AI assistants (at work, home, etc.) introduces a new set of inference-time privacy risks: LLMs are fed different types of information from multiple sources in their inputs and are expected to reason about what to share in their outputs, for what purpose and with whom, within a given context. In this work, we draw attention to the highly critical yet overlooked notion of contextual privacy by proposing ConfAIde, a benchmark designed to identify critical weaknesses in the privacy reasoning capabilities of instruction-tuned LLMs. Our experiments show that even the most capable models such as GPT-4 and ChatGPT reveal private information in contexts that humans would not, 39% and 57% of the time, respectively. This leakage persists even when we employ privacy-inducing prompts or chain-of-thought reasoning. Our work underscores the immediate need to explore novel inference-time privacy-preserving approaches, based on reasoning and theory of mind.
hub tools
citation-role summary
background 3
method 1
citation-polarity summary
representative citing papers
PrincipalBench exposes a sharp split in frontier LLMs between selective and over-refusing behavior on multi-party loyalty, with prompt scaffolding and KL distillation reducing harm rates but only along an existing leak/over-refusal trade-off.
OCELOT recasts agent privacy as posterior-risk control and implements Witness-Verified Declassification to authorize the least-disclosing useful release under a sink-trust-weighted min-entropy budget.
Anchored Privacy Drifting (APD) replaces privacy-sensitive visual elements with semantically equivalent alternatives while anchoring context, evaluated on the new AdaptShield benchmark with reported gains of 10.4% and 8.5% across four MLLM families.
Frontier LLMs leak prompted secret information thematically in generated stories at rates up to 79% above chance in binary discrimination tests, even when told to hide it, with leakage scaling by model size and vanishing for short-form outputs.
Users show curiosity over concern toward LLM inferences of personal information, with acceptability depending on context, alignment with expectations, and who uses the inferences rather than just the content.
CalBench is a new benchmark for multi-agent LLM calendar scheduling that measures task success, excess cost, communication efficiency, burden fairness, and privacy leakage under private information constraints.
CAMP formalizes Cumulative PII Exposure and uses a session registry, co-occurrence graph, and CPE score to trigger retroactive masking in multi-turn LLM conversations, neutralizing re-identifiable profiles in synthetic tests while keeping utility intact.
Memory-equipped LLM agents exhibit increasing safety violation rates as memory accumulates across independent tasks, termed temporal memory contamination, detected via a new trigger-probe protocol.
PrivScope enforces task-scoped disclosure at the local-cloud boundary in hybrid agents, eliminating profile leakage and halving re-identification risk on medical workflows while preserving task success.
AgentCollabBench shows that multi-agent reliability is limited by communication topology, with converging-DAG nodes causing synthesis bottlenecks that discard constraints and explain 7-40% of information loss variance.
Vision-language models exhibit perceptual fragility and fail to consistently respect privacy constraints when operating in simulated physical environments, with performance declining in cluttered scenes and under conflicting commands.
A data-centric survey finds that only information-flow control covers compositional and cross-session leakage in LLM agents and that no single benchmark tests an agent across all its data surfaces under one policy.
The paper defines four governance primitives for shared memory in multi-agent LLM fleets and reports evaluation results from a live production service.
The paper reviews the background, technology, applications, limitations, and future directions of OpenAI's Sora text-to-video generative model based on public information.
citing papers explorer
-
The Impact of Security and Privacy Controls on Users' Emotional Engagement with Generative AI Chatbots
In a vignette study of 354 U.S. participants, deletion-based privacy controls outperformed all other controls in increasing willingness to engage with GenAI chatbots for emotional support, while technically complex controls like local-only processing and model training opt-outs reduced engagement.
-
Whose Side Is Your Agent On? Multi-Party Principal Loyalty in LLM Agents
PrincipalBench exposes a sharp split in frontier LLMs between selective and over-refusing behavior on multi-party loyalty, with prompt scaffolding and KL distillation reducing harm rates but only along an existing leak/over-refusal trade-off.
-
OCELOT: Inference-Leakage Budgets for Privacy-Preserving LLM Agents
OCELOT recasts agent privacy as posterior-risk control and implements Witness-Verified Declassification to authorize the least-disclosing useful release under a sink-trust-weighted min-entropy budget.
-
Seeing Without Exposing: Adaptive Privacy Control for Open-World, Context-Hungry MLLMs
Anchored Privacy Drifting (APD) replaces privacy-sensitive visual elements with semantically equivalent alternatives while anchoring context, evaluated on the new AdaptShield benchmark with reported gains of 10.4% and 8.5% across four MLLM families.
-
Can You Keep a Secret? Involuntary Information Leakage in Language Model Writing
Frontier LLMs leak prompted secret information thematically in generated stories at rates up to 79% above chance in binary discrimination tests, even when told to hide it, with leakage scaling by model size and vanishing for short-form outputs.
-
When Are LLM Inferences Acceptable? User Reactions and Control Preferences for Inferred Personal Information
Users show curiosity over concern toward LLM inferences of personal information, with acceptability depending on context, alignment with expectations, and who uses the inferences rather than just the content.
-
CalBench: Evaluating Coordination-Privacy Trade-offs in Multi-Agent LLMs
CalBench is a new benchmark for multi-agent LLM calendar scheduling that measures task success, excess cost, communication efficiency, burden fairness, and privacy leakage under private information constraints.
-
CAMP: Cumulative Agentic Masking and Pruning for Privacy Protection in Multi-Turn LLM Conversations
CAMP formalizes Cumulative PII Exposure and uses a session registry, co-occurrence graph, and CPE score to trigger retroactive masking in multi-turn LLM conversations, neutralizing re-identifiable profiles in synthetic tests while keeping utility intact.
-
Remembering More, Risking More: Longitudinal Safety Risks in Memory-Equipped LLM Agents
Memory-equipped LLM agents exhibit increasing safety violation rates as memory accumulates across independent tasks, termed temporal memory contamination, detected via a new trigger-probe protocol.
-
PrivScope: Task-scoped Disclosure Control for Hybrid Agentic Systems
PrivScope enforces task-scoped disclosure at the local-cloud boundary in hybrid agents, eliminating profile leakage and halving re-identification risk on medical workflows while preserving task success.
-
AgentCollabBench: Diagnosing When Good Agents Make Bad Collaborators
AgentCollabBench shows that multi-agent reliability is limited by communication topology, with converging-DAG nodes causing synthesis bottlenecks that discard constraints and explain 7-40% of information loss variance.
-
How Far Are VLMs from Privacy Awareness in the Physical World? An Empirical Study
Vision-language models exhibit perceptual fragility and fail to consistently respect privacy constraints when operating in simulated physical environments, with performance declining in cluttered scenes and under conflicting commands.
-
Agents That Know Too Much: A Data-Centric Survey of Privacy in LLM Agents
A data-centric survey finds that only information-flow control covers compositional and cross-session leakage in LLM agents and that no single benchmark tests an agent across all its data surfaces under one policy.
-
Governed Shared Memory for Multi-Agent LLM Systems
The paper defines four governance primitives for shared memory in multi-agent LLM fleets and reports evaluation results from a live production service.
-
Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models
The paper reviews the background, technology, applications, limitations, and future directions of OpenAI's Sora text-to-video generative model based on public information.