SNAPRAG extracts RAG knowledge bases with fewer black-box queries
The attack mutates tokens without an external LLM and, the authors say, needs up to 6.77 times fewer queries to extract half a knowledge base.
Stories per day · 30 days
2 RSS
Relatedadvbench1adversarial-attacks1adversarial-queries1detection-evasion1drl1evasion1false-data-injection1fingerprinting1harmbench1image-classifiers1jailbreak1kgw1knowledge-extraction1llm-agents1llm-safety1llm-watermark1markov-chains1model-attribution1
28 stories, newest first · one card per story
Stories tagged #black-box
ScrambleMark evades speech watermarking via harmonic scrambling
ScrambleMark achieves 97% evasion on commercial platforms while preserving speech naturalness at 4.2/5.
Black-box fake V2I reports delay DRL traffic signals
Simply reporting the maximum feasible queue was as damaging as a trained surrogate; random inflation was not.
STROKE uses sparse rectangular occlusions for black-box image attacks
The score-based method reports high success at 1–2% area with low query counts and shorter wall-clock time than matched patch attacks.
AgentTrace fingerprints black-box LLM agents from execution traces
Markov models of repair sequences reach 96.0% attribution accuracy and separate Llama-3 8B from 70B under spoofing.
Black-box attack defeats statistical LLM watermarks including SynthID
The attack drops detection to at most 0.05 on seven schemes and three models without keys or weights.
ECHO jailbreak search scores reply harmfulness instead of compliance prefixes
Authors report up to 0.898 StrongREJECT and 89.23% HarmBench success on five open-weight models; code is public.
Hide and Seek: HIDE Concepts to SEEK explanations for black box models
Hide and Seek attributes concepts in black-box vision models without weights or labeled concept datasets.
Black-Box Adversarial Patch Attacks on VLAs via Ancestor VLM Exploitation
Adversarial patches optimized only on ancestor VLMs substantially cut success rates of deployed vision-language-action models.
CorrectGuard: Eyes-Off Correctness Estimation for Black-Box Security Guardrails
CorrectGuard estimates black-box AI guardrail correctness without inspecting user input or guardrail internals, boosting error-detection accuracy up to 25 points.
SteerProbe: Learning to Bypass Safety Steering in Vision-Language Models
SteerProbe bypasses vision-language model safety steering, lifting average harmful responses from 7.43% to 18.36%.
Dagger: Decoupling-based Model Stealing Attack against Graph Neural Networks
Dagger steals GNN models from hard-label APIs with higher fidelity and far fewer queries.
Controlled Decoding Attacks on Black-Box LLMs
A text-only controlled-decoding method jailbreaks black-box LLMs by reconstructing and steering token distributions at a few positions.
Practical Secrets Extraction against Black-box LLMs
Researchers recover memorized credentials from black-box commercial LLMs, including OpenAI and Claude Code.
Controlled Decoding Attacks on Black-Box LLMs
Researchers propose a text-only black-box LLM jailbreak that reconstructs and selectively steers token distributions.
DeMark: A Query-Free Black-Box Attack for Quality-Preserving Audio Watermark Removal
DeMark removes learning-based audio watermarks query-free, reaching 0.92–0.96 success while preserving speech quality.
AnswerMap: Faithful Spatial Interpretability of VLMs from Answer Posteriors
AnswerMap builds black-box spatial rationales for vision-language models from output answer posteriors.
Statistical attribute alignment for black-box generative AI via output post-processing
Researchers propose query-efficient post-processing to align generative AI attributes with a target distribution.
Region-Level Black-Box Defense Against Stealthy Embedding-Space Backdoors in CLIP
CLIPGuard localizes and inpaints CLIP embedding backdoors, cutting attack success to about 1% in black-box tests.
JevAdvBench: A Benchmark and Black-Box Attacks for Reinforcement Learning for Calibrated Decisions Models
JevAdvBench shows appended opinions flip 12.1% of Jev decisions and sink many below review confidence.
Breaking the Black Box: Byte-Level Boundary Inference of Real-World Antivirus Systems
AVHunter infers byte-level decision boundaries of 11 antivirus products, exposing compact regions that enable evasion.
Extracting CNNs in the Unknown-Architecture and Feedback-Agnostic Setting
Researchers show black-box attacks can recover a CNN's unknown architecture and its parameters.
GuidedRay: Diversity-Guided Direction Discovery for Targeted Hard-Label Black-Box Attacks
GuidedRay finds targeted adversarial directions for hard-label black-box attacks using fewer queries than prior methods.
Feedback Coding Enables Inference-Time Covert Agentic Communication
Researchers show black-box LLMs can hide messages in generated text with near-zero decoding error.