REVIEW 15 cited by
Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
We demonstrate how images and sounds can be used for indirect prompt and instruction injection in multi-modal LLMs. An attacker generates an adversarial perturbation corresponding to the prompt and blends it into an image or audio recording. When the user asks the (unmodified, benign) model about the perturbed image or audio, the perturbation steers the model to output the attacker-chosen text and/or make the subsequent dialog follow the attacker's instruction. We illustrate this attack with several proof-of-concept examples targeting LLaVa and PandaGPT.
Forward citations
Cited by 15 Pith papers
-
Agent Security Needs Redefinition through a Holistic Framework
Agent security should be redefined around four contextual authorization properties instead of the content of the action performed.
-
Do Agents Dream of False Memories? Black-box Visual Attacks on Long-term Memory in Multimodal AI Agents
Lucid shows that imperceptible image perturbations can make multimodal agents misremember past events with 61.6% poisoning and 58.4% injection success.
-
The Self-Correction Illusion: Role Relabeling Gates Explicit Error Flagging in Large Language Models
Relabeling an identical erroneous claim from the model's own thought role to an external chat role increases explicit correction rates by 23-93 percentage points across 13 model-domain cells, indicating a chat-templat...
-
A Synonymous Variational Perspective on the Rate-Distortion-Perception Tradeoff
Synset-based reconstruction and synonymous variational inference are claimed to derive the distributional divergence in RDP and unify it with classical rate-distortion theory.
-
Prevalence of Security and Privacy Risk-Inducing Usage of AI-based Conversational Agents
Roughly a third of UK adults use AI chatbots weekly, and among them a substantial minority upload untrusted content, connect bots to other programs, share sensitive data, or attempt jailbreaks.
-
Invitation Is All You Need! Promptware Attacks Against LLM-Powered Assistants in Production Are Practical and Dangerous
Malicious calendar invites and emails can poison Gemini's context, enabling data exfiltration, app control, and physical-world actions.
-
Secure Tug-of-War (SecTOW): Iterative Defense-Attack Training with Reinforcement Learning for Multimodal Model Security
An iterative attacker-defender reinforcement learning method that makes a multimodal LLM refuse more jailbreak prompts without over-refusing ordinary queries.
-
Normative Conflicts and Shallow AI Alignment
A philosophical argument that preference fine-tuning produces shallow alignment, plus a new 'thought injection' attack that exploits reasoning traces in LLMs.
-
Con Instruction: Universal Jailbreaking of Multimodal Large Language Models via Non-Textual Modalities
Con Instruction embeds harmful textual instructions into adversarial images or audio by aligning their representations, achieving successful jailbreaks on several vision- and audio-language models.
-
Devil in the Lens: Analyzing and Defending Physical Prompt Injection Against Vision-Language Models on Wearable Devices
Physical scene text can inject prompts into wearable VLMs, hijacking decisions and content with high success rates across six threat scenarios, partially mitigated by OCR masking and token-drift defenses.
-
Layer-Wise Perturbations via Sparse Autoencoders for Adversarial Text Generation
Sparse autoencoder activation perturbation (SFPF) applied on top of existing jailbreak prompts raises attack success rate on Qwen3-32B, but with no defense evaluation and weak reproducibility.
-
The Boy Who Cried Wolf: Adversarial Misclassification of Safe Inputs as Unsafe in Multimodal Guardrails
Adversarial images aligned with the latent distribution of unsafe content can force multimodal guard models to falsely reject safe user requests with up to 84% success.
-
Seven Security Challenges in Cross-domain Multi-agent LLM Systems
The paper proposes that cross-domain multi-agent LLM systems face seven novel security challenge categories that existing single-domain defenses do not cover.
-
A Survey of Safety on Large Vision-Language Models: Attacks, Defenses and Evaluations
A survey of LVLM safety that adds a lifecycle taxonomy and new benchmark results showing Janus-Pro-7B has weaker safety than several open-source LVLMs.
-
Prompt Injection 2.0: Hybrid AI Threats
A structured taxonomy of hybrid prompt injection attacks shows how XSS, CSRF, and SQL injection vectors converge with LLM manipulation to bypass traditional controls.
Discussion (0). Sign in to comment.