Kerkoporta
Structured-output backdoors inserted via Kerkoporta reach success rates of up to 96.2% in reasoning coding agents.
Injecting Structured-Output Backdoors in Reasoning Coding Agents
AI Research Radar
Tuesday, November 3: 7 frontier papers out of 954 read.
Frontier
Papers rated as moving the field
How we chooseMAC-Bench
One Waveform, Two Stories
On SAC-Bench, Event Identity accuracy is 24.8--46.5 percentage points lower under conflicting claims than compatible ones.
DexArena
DexArena introduces 10 long-horizon and 30 stage-level bimanual dexterous manipulation tasks, with baselines revealing substantial challenges in precision, generalization, and skill composition.
Are Fake Citations a Form of Reward Hacking? A Pre-Registered Study of Post-Training and Citation Cues
Asking for citations raises fabrication sharply in post-trained checkpoints but not base ones, a 47.8-point interaction.
Rewarded for Not Getting Caught
Reward-based alignment training can only encode norms as prices charged when violations are noticed, favoring conditional compliance.
AgiWorld
AgiWorld builds persistent worlds from a single image and free-form instruction, achieving state-of-the-art performance on world revisiting and exploration benchmarks.
This week in research
Oct 28 – Nov 3: 269 cleared, 19 frontier
- Security98◆ 3
- Agents36◆ 2
- Safety35◆ 5
- Society23
- Robotics20◆ 2
- 4 moreTraining, ML, Science, Reasoning16◆ 4
- Vision13◆ 2
- Inference10
- Language9◆ 1
- Security98◆ 3
- 8 moreRobotics, Vision, Inference, Language, Training, ML, Science, Reasoning68◆ 9
- Agents36◆ 2
- Safety35◆ 5
- Society23
Size: cleared papers◆ and tint: frontier papers9 papers not yet sorted into a field
Also cleared
106 papers, by field
Security & privacy33
- MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI AgentsAlibabaCode
- SearchInject: Retrieval Exposure Control for Indirect Prompt Injection in Language-Model SearchGoogle
- HideOut - User Protection Against Localization with Throughput Guarantees in Mobile Networks
- Humanized but Still True? Auditing Facts and Detection in AI-Text Humanizers
- When appearance and motion disagree: evaluating class--motion consistency against disguised uncrewed surface vessels
All 33 papers in Security & privacy
- Threat Modeling Intimate Partner Abuse in Agentic Systems
- The Long Con on the Web: Earn-then-Exploit Fraud Attacks against Coding Agents
- Constructing Budgeted KGW Evasion through Token-Event Evidence Replacement
- AgentCap: Capability-Based Runtime Authorization for Secure LLM Agent Execution
- Inherited or Learned? Backdoor Selectivity in Distilled Language Models
- Illusion: Safeguarding Topology Confidentiality in LLM Multi-Agent Systems via Behavior-Preserving Obfuscation
- The Blind Spots of Detection: Adversarial Perturbation Attacks against Watchdogs in Distributed LLM Inference
- Do GEO Gains Survive the Live Web? Evaluating In-the-Wild Visibility Gains of Generative Engine Optimization via LIVE-GEO
- Context-Shielded Backdoor Attack against Large Vision-Language Models
- J-Anchor: Defending Against Prefill Attacks via J-Space Selective Refusal Restoration
- Sensitivity Inflation from Cross-Record Precision Routing in Private Quantized Fine-Tuning
- Label-Free Localization and Test-Time Suppression of CLIP's Typographic Circuits
- Launder-Bench: Can Repository-Level Agentic Code Laundering Be Detected?
- Beyond Prompt Injection: The Tool Update Gap in LLM Agent Security
- SANAI: Secure ECC Distributed Keyed Parity Embedded in Weights for Protecting Neural Networks against Bit-Flip Attacks
- Black-Box Adversarial Example Attacks on Machine Learning based Election Systems
- Detect, Repair and Audit: A Hierarchical Multi-Agent Defense Against Adversarial Attacks in Energy Forecasting
- Accessed, never disclosed: Silent access to protected health information by medical artificial intelligence agents
- Networked Synthetic Social Engineering on the Web
- Probing Local Conditional Fit for Diffusion Model Membership Inference
- StateFence: Task-State Protection Against Evidence-Chain Contamination in Blue-Team LLM Agents
- Where Does MCP Security End? A Three-Tier Mitigation Mapping Framework for the OWASP MCP Top 10
- ReplayFactor: Replay-Aware Dual-Attribute Learning for Practical Speech Deepfake Detection
- The Customer Never Sees the Goal: Measuring and Detecting Vulnerability-Targeted Manipulation in LLM Assistants
- ATrust: Adaptive Trust Boundary Expansion for Securing LLM Agents
- Evaluating autonomous self-healing against persistent cyber-attacks
- IAMRAGBENCH: Benchmarking Authorization Expansions in LLM-Generated AWS IAM Policies under RAG Poisoning
- MCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study
Agents15
- Decentralized Multi-Agent Coordination and Communication in Shared CodebasesNVIDIA
- RECEIPT: Measuring and Repairing the Return Assumption of LLM Agents in Web Service OperationsTencent
- ParetoSaddler: Pareto-Guided Optimization of Agent Harnesses for Quality and CostMicrosoft
- Axion 1.0: Open Agent Foundation Models, SimplifiedByteDance
- VisCodeBench: a new benchmark dataset for Text-to-Vis reflecting real-world practiceCode
All 15 papers in Agents
- Fidelity Collapses with Table Size, Not Turn Count: Delivery Architecture in a Deployed Tool-Using Agent
- Do Coding Agents Catch Their Own Bugs? (They Don't.)
- Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces
- Democratizing Robot-Use Agents: A Unified Harness with Open Models
- Trustworthy Execution of Industrial Agents
- CAREFUL-Bench: Effect-Aware, Judge-Free Evaluation of Multi-Turn LLM Function Calling
- DR Pixel: Page-Native Deep Research with Vision-Language Models
- DeliverBench: Benchmarking AI Agents on Delivering Realistic Multi-Stage Professional Projects
- Optimize the Harness, Not the Model: Fitting a Minimal Agent to a Self-Hostable LLM for Investigative Analysis
- An Agentic Framework for Automated Power System Studies and Grid Connection Assessment
Safety & alignment12
- PGuard: A Native, Streaming, and Plug-in Guardrail for Multimodal Large Language ModelsAlibaba
- Dependence-Aware Risk Assessment and Optimal Design of AI-Agent Safeguard Systems
- Circuit Oracle: Automating Attribution Graph Analysis via Natural-Language Queries
- What Does a Safety Label Measure? Auditing Situational-Safety Benchmarks for Multimodal LLMs
- Satisfied, Not Solving: Wireheading of Bounded Homeostatic Rewards in Reinforcement Learning
All 12 papers in Safety & alignment
- Large language models underrepresent malice and overreport virtue
- Multi-Agent Collusion and the J-Space
- A balloon effect: ban one channel and a hidden trait finds another
- Cumulative and Audited Guardrail Tool Evolution for LLM Agent Safety
- EEFA: A Framework for the Ethical Evaluation of AI Agents
- Predictive Safety Guardrails Do Not Always Make LLM Agents Safer
- VeriHarm-Bench: Measuring Harmful Success in LLM Coding Agents Verifiably
Robotics & embodied AI11
- RoboColosseum: Evaluating Physical Intelligence at Scale In-The-WildAllenAI, Hugging Face
- Auriga Technical Report: A Task-Agnostic Agent Harness for Zero-Shot Embodied ManipulationCode
- KlaskTron: An Open-Source Platform for Physical Adversarial Multi-Agent RLCode
- Does Validation Loss Predict Real-Robot Success? Evidence from Four Imitation-Learning Policies on a Low-Cost Arm
- Leap-VLA: High-Rate Robot Control from a Low-Rate VLM Backbone
All 11 papers in Robotics & embodied AI
- Route-Guided CBF-QP Repair for Safe Vision-Language-Action Models
- When Delay Meets Drift: Physical Safety of Robot Foundation Models Is a Cross-Layer Problem
- W2A: Distilling World Priors into Action Models for Real-Time Robotic Control
- KlaskTron-IL: A Contact-Rich Adversarial Imitation Learning Benchmark with a Validated Sim-to-Real Bridge
- Rollout then Recycle: 3,718 Labelled Real-Robot Episodes from Benchmarking 7 Policies on 12 Tasks
- Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation
AI & society9
- Measuring User Success in Randomized Search Experiments: Interpreting Good Abandonment in Generative AI SearchMicrosoft
- Not All Minorities Are Equal: Fairness Outcomes in Private Image Generation
- When AI Agents Act on Our Behalf: A Taxonomy of Human-Agent Interaction Failure
- Queering the Audits: Community-Based Auditing of AI Harms to Queer Communities
- Precedent Labels and Facts Shift LLM Approval Decisions in Opposite Directions
All 9 papers in AI & society
- Same Sentence, Different Construct: Identity-Swapped Rationales Reveal Measurement Non-Invariance in LLM Hate Speech Classifiers
- From Language to Country to Valence: A Multilingual Audit of Stereotype Bias Across LLMs
- LLMs Cannot Be Neutral: Gender Sensitivity Across Candidates, Evaluators, and Models
- ComplianceAtlas: A Taxonomy-Driven Agentic AI Framework for Automated AI Act Gap Analysis and Remediation
Inference & efficiency6
- Jet-QSparse: Accelerating Agentic LLM Inference with 4-Bit Sparse AttentionNVIDIA
- Scute: Exact Batched Serving for PEFT Adapters Beyond LoRA
- Serving Agentic Robot Fleets: Workload Characterization, KV-Cache Thrashing, and Admission Control
- Seeing What Matters, Verifying What Pays: Load-Adaptive Speculative Decoding for Vision-Language Models
- A Compute Center Anyone Can Join: A Peer-to-Peer System for AI Inference
- The SIMPLE Law for Energy Cost of LLM Inference
Language & speech5
- GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled DataAlibaba
- To Jev or Not to Jev: When Can System-1 Decision Models Replace LLM Judges?
- Federated Vector Retrieval across Heterogeneous VectorDBs
- When the Text Is Silent: Memory Shortcuts and the Validity of LLM Annotation under Evidence Insufficiency
- COVE: A Task-Adaptive, Code-Based Ensemble for Proactive Verification of LLM Outputs
Vision & multimodal3
Science & health2
Machine learning1
Not yet sorted9
- On the Framing Effect of AI Red-Teaming Instructions: Evidence from Prior Exercises and Controlled SimulationsGoogle
- XLA from the Trenches: Correctness on TPU Workloads at ScaleGoogle
- SpecPilot: Autonomous Algorithm Discovery and Systems Optimization for Speculative DecodingNVIDIA
- CISA: Retrofitting Dense Attention Models with Chunk-Indexed Sparse AttentionNVIDIA
- A Novel Federated Learning Framework for Defending Against Gradient Inversion AttacksCode