adversarial ml
30 articles about adversarial ml in AI news
New Research Proposes FilterRAG and ML-FilterRAG to Defend Against Knowledge Poisoning Attacks in RAG Systems
Researchers propose two novel defense methods, FilterRAG and ML-FilterRAG, to mitigate 'PoisonedRAG' attacks where adversaries inject malicious texts into a knowledge source to manipulate an LLM's output. The defenses identify and filter adversarial content, maintaining performance close to clean RAG systems.
VMLOps Publishes NLP Engineer System Design Interview Guide
VMLOps has published 'The NLP Engineer's System Design Interview Guide,' a detailed resource covering architecture, scaling, and trade-offs for real-world NLP systems. It provides a structured framework for both interviewers and candidates.
DEAF Benchmark Reveals Audio MLLMs Rely on Text, Not Sound, Scoring Below 50% on Acoustic Faithfulness
Researchers introduce DEAF, a 2,700-stimulus benchmark testing Audio MLLMs' acoustic processing. Evaluation of seven models shows a consistent pattern of text dominance, with models scoring below 50% on acoustic faithfulness metrics.
AI Teaches Itself to See: Adversarial Self-Play Forges Unbreakable Vision Models
Researchers propose AOT, a revolutionary self-play framework where AI models generate their own adversarial training data through competitive image manipulation. This approach overcomes the limitations of finite datasets to create multimodal models with unprecedented perceptual robustness.
New Paper Coins 'Curation Debt' — Benchmarks Measure Data Leakage, Not Capability
New paper coins 'curation debt' — benchmarks like MMLU measure data leakage, not capability. Proposes adversarial dynamic benchmarks.
Microsoft RAMPART Brings Pytest-Based Safety Testing to AI Agents
Microsoft's RAMPART brings pytest-native safety testing to AI agents, covering adversarial attacks and benign failures, addressing a critical gap in agent development.
How a Custom Multimodal Transformer Beat a Fine-Tuned LLM for Attribute
LeBonCoin's ML team built a custom late-fusion transformer that uses pre-computed visual embeddings and character n-gram text vectors to predict ad attributes. It outperformed a fine-tuned VLM while running on CPU with sub-200ms latency, offering calibrated probabilities and 15-minute retraining cycles.
Why Production AI Needs More Than Benchmark Scores
The article argues that high benchmark scores are insufficient for production AI success, highlighting the need for robust MLOps practices, monitoring, and real-world testing—critical for retail applications.
POTEMKIN Framework Exposes Critical Trust Gap in Agentic AI Tools
A new paper formalizes Adversarial Environmental Injection (AEI), a threat model where compromised tools deceive AI agents. The POTEMKIN testing harness found agents are evaluated for performance, not skepticism, creating a critical trust gap.
Uni-SafeBench Study: Unified Multimodal Models Show 30-50% Higher Safety Failure Rates Than Specialized Counterparts
Researchers introduced Uni-SafeBench, a benchmark showing that Unified Multimodal Large Models (UMLMs) suffer a significant safety degradation compared to specialized models, with open-source versions showing the highest failure rates.
MAIL Network: A Breakthrough in Efficient and Robust Multimodal Medical AI
Researchers have developed MAIL and Robust-MAIL networks that overcome key limitations in multimodal medical imaging analysis, achieving up to 9.34% performance gains while reducing computational costs by 78.3% and enhancing adversarial robustness.
Decoy Font Tricks AI Vision Models With Dual-Layer Glyphs
Mixfont's Decoy Font hides text from AI vision models by layering two characters into one glyph, exploiting a tokenization blind spot in ChatGPT and Gemini.
gdb: Benchmarks Saturate Too Fast for Reliable AI Progress Tracking
@gdb notes benchmarks saturate quickly. This undermines AI progress tracking and may force shift to dynamic evaluations.
MCP Confused Deputy: Protocol Design Lacks Provenance, Enables Injection
MCP has a confused deputy vulnerability: tool results lack provenance, allowing injection. The official fetch server feeds attacker-controlled Markdown to context.
Claude Code Digest — Jul 10–Jul 13
Claude Code is crossing the line from “assistant” to “agent runtime”: the winning teams are the ones adding verification, hooks, and policy gates instead of trusting the model.
Databricks Tests Coding Agents on Its Own Codebase
Databricks benchmarked coding agents on its own polyglot codebase. GLM-5.2 matched top closed models, a minimal harness halved costs, and cheaper-per-token models cost more per task.
Claude Code Steganography Flagged Chinese Users; Anthropic Rolls Back
Anthropic's Claude Code 2.1.91 used steganography to detect Chinese users. After Reddit exposure, Anthropic rolled back the feature, calling it an experiment against model distillation.
SingGuard: Runtime Guardrails for Multimodal AI Treat Safety as Input
SingGuard treats safety rules as runtime inputs for multimodal AI, achieving SOTA across 6 families and 35 datasets via fast/slow reasoning.
Stop Testing Skills Once: Use Caliper's pass@k to Measure What Actually
Caliper is a lightweight harness that runs Claude Code skills k times, scores them with pass@k, and compares against a no-skill baseline so you know if your skill actually helps.
Caliper: Run Your Claude Code Skills k Times and Get a pass@k Score That
Caliper gives Claude Code users a pass@k reliability score for skills, with a baseline delta showing if the skill beats the base agent. Install via pipx or npx.
RIFT-Bench Tests 45 Agentic Systems With Dynamic Red-Teaming
RIFT-Bench evaluates 45 agentic AI systems via a graph-driven two-phase pipeline, enabling unified security comparison across heterogeneous architectures.
The AI benchmark gap has collapsed: top 10 labs now separated by just 44 Elo points
Chatbot Arena Elo scores and Artificial Analysis data confirm that the top 10 AI labs are now clustered within 44 Elo points — the narrowest spread on record. Stanford HAI's 2026 AI Index corroborates the trend: leading frontier models are separated by as little as 3 percentage points on most benchm
SMAC-Talk: StarCraft Benchmark Tests LLM Agents Against Deceptive Allies
SMAC-Talk extends StarCraft Multi-Agent Challenge with natural language communication, testing LLM agents against deceptive allies. Qwen3.5 models benchmarked; no model exceeds 72% win rate.
Ontology-Grounded AI Agent Testing Hits 48.3% Regulatory Coverage vs.
Ontology-grounded AI agent testing achieves 48.3% regulatory coverage vs. 33.1% baseline in 1800-scenario pilot. Coverage advantage over RAG not robust after Bonferroni correction.
Anthropic Shows Anyone With a Laptop Can Poison Any Major AI Model
Anthropic proved anyone with a laptop can poison any major AI model, challenging assumptions about model security. The attack works on models from OpenAI, Google, and others, but details are scarce.
Embedding distance predicts VLM typographic attack success (r=-0.93)
A new study shows that embedding distance between image text and harmful prompt strongly predicts attack success rate (r=-0.71 to -0.93). The researchers introduce CWA-SSA optimization to recover readability and bypass safety alignment without model access.
SharpAP: New Attack Method Makes Recommender System Poisoning More
Researchers propose SharpAP, a poisoning attack that uses sharpness-aware minimization to generate fake user profiles that transfer better between different recommender system models, posing a more realistic threat.
AI Hiring Tool Rejects Same Resume Based on Name Change
Researchers sent identical resumes to an AI hiring tool, changing only the name. One version was rejected, revealing systemic bias in automated hiring systems.
Continuous Semantic Caching
Researchers propose a theory-grounded semantic caching system that treats user queries as points in a continuous embedding space, using dynamic ε-net discretization and kernel ridge regression to cut inference costs and latency without switching overhead.
DNL Method Finds 2 Bits That Crash ResNet-50, Qwen3-30B
Researchers introduced Deep Neural Lesion (DNL), a method to find critical parameters. Flipping just two sign bits reduced ResNet-50 accuracy by 99.8% and Qwen3-30B reasoning to 0%.