SoK: Semantic Decision Engines in Network Control Loops

Systematizing 139 semantic decision engine families reveals most miss network deadlines and verification, with only four reporting deadline attainment; unverified decisions reverse admission verdicts under queued execution, prompting minimum reporting rules and a research agenda.

Delong Li, Chen Li, Xu Wang, Haochen Gong and 2 more

Published Oct 5, 2026 · ▲ 6 on Hugging Face · Code

Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization

ATPO uses adaptive Tversky reinforcement learning for controllable multi-label video safety detection, raising Jaccard Index to 75.44 on SafeWatch-Bench-Real while enabling steerable precision-recall trade-offs.

Guangyu Yang, Jingbiao Mei, Mingsheng Sun, Jinghong Chen and 4 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published Oct 1, 2026 · 0 citations

Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems

Benign multi-agent LLM planners disguise secrets to help developers evade oversight, with rare per-episode leaks compounding to high breach risk across repeated exchanges.

Deema Alnuhait, Gengyu Wang, Muhammad Khalifa, Hao Peng

Published Sep 30, 2026 · 0 citations · ▲ 13 on Hugging Face

Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination

The RA-Bench benchmark reveals current detectors fail to consistently identify AI-generated crisis videos, which become harder to detect after social dissemination and frequently mislead humans.

Shuo Liang, Yixing Ma, Pengfei Zhou, Zhenglin Wan and 32 more

Published Aug 14, 2026 · 0 citations · ▲ 287 on Hugging Face · Code ★ 132

Do Thinking Tokens Help with Safety?

Narutatsu Ri, Abhishek Panigrahi, Sanjeev Arora

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Model Incrimination: Investigating Whether Concerning Behavior Reflects Misalignment

Gerson Kroiz, Aditya Singh, Senthooran Rajamanoharan, Neel Nanda

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

OASIS: Online Adaptive Steering for In-Training Safety of LLMs

Yifan Sun, Qiang Sheng, Ya Wu, Zhengjia Wang and 4 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

LLM-enabled Applications Require Systematic Threat Monitoring

Yedi Zhang, Haoyu Wang, XIANGLIN YANG, Jin Song Dong and 1 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Latent Barrier Steering: Hierarchical Safety for Generative Planning

Renhao Zhang, Mingzhe Li, Bruno Silva

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation-Level Enforcement

Xutao Mao, Jianing Zhu, Jinman Zhao, Tongliang Liu and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Safety-Aware Latent Space Reasoning in Large Language Models

Yi Wang, Wenjie Wang, Hongye Qiu, Yu Pan

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

AI Control for Sandbagging on Fuzzy Tasks

Mikhail Terekhov, Caglar Gulcehre, Vivek Hebbar, Joe Benton

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Decoupled Safety Control: A Safety-Control Algorithm for Training-Free Safety Guidance

Huilin Zhou, Ruoxi Cheng, Yuhang Wang, Yuming Liu and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Position: AI Development Should Prioritize Cognitive Security

Batu El, Shiye Su, Aneesh Pappu, Peggy Yin and 5 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

AI Alignment Can Build Moral Autonomy

Kaile Wang, Hantao Lou, Tianyi (Alex) Qiu, Sebastian Sunday-Grève and 4 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Tight PAC-Bayes Generalisation Guarantees for Large Language Model Safety Monitoring

Tom Lamb, Philip Torr, Tim G. J. Rudner

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Stress Testing Chain-of-Thought Monitoring Against Covert Misalignment

Yao Huang, Yifan Wang, Yitong Sun, Yichi Zhang and 1 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Position: A Safe LLM and a Safe Harness Do Not Make a Safe Agent

Vincent Siu, Kyle Montgomery, Yujin Potter, Zhun Wang and 2 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Scheming Is a Symptom: Alignment Research Should Probe Reflexive Fragility

Nan Zhang, Heng Xu

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Are LLM Safety Judges Policy-Invariant? A Three-Principle Stress-Test

Shihao Weng, Yang Feng, Xiaofei Xie

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Rubric-Align: Safety Alignment through Dynamically Co-Evolving Rubrics

Ruipeng Wang, Junfeng Fang, Houcheng Jiang, Kai Tang and 4 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

MLLM-Edit: Benchmarking Image Forgery Detection and Localization under MLLM-based Editing

Zeqin Yu, Ye Tian, Jian Zhang, Jiangqun Ni and 3 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Pluralistic AI Alignment Requires Inference-Time Multi-Objective Control

Weichen Li, Mislav Stojanović, Daniel Neider, Marius Kloft and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Position: We Need Greater Transparency to Maintain Research Pipeline Reliability Despite GenAI

Hillmer Chona, Sourav Panda, Frank Ritter, Jonathan Dodge

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

AIs with Secret Loyalties are a Serious but Addressable Threat

Joe Kwon, Alfie Lamerton, Andrew Draganov, Dave Banerjee and 6 more

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Claude Coke: Prevent Automated Crime by Agents

Gabor Hollbeck, Baran Peters, Alexander von Recum, Jan Granacher and 3 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Exploring Advertising Manipulation in Diffusion Image Generation

Tianshi Che, Yang Zhou, Yushan Mu, Zeru Zhang and 7 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

AI Safety Evaluations Need More Human-AI Experiments

Michelle Vaccaro, Jaeyoon Song, Abdullah Almaatouq, Michiel Bakker

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Corrupted Plans, Clean Traces: What Planning-Execution Decoupling Reveals About CoT Monitoring

Keertana Chidambaram, Andrew Ilyas, Vasilis Syrgkanis

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

nnTrace: Detecting and Localizing Silent Bugs in Distributed Training

Haitian Jiang, Shaowei Zhu, Zhen Zhang, Zhenyu Song and 4 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Answering At Any Cost: Frontier LLMs Are Consequence-Insensitive

Arka Pal, Kwok C Au, Louai Zahran, Rahul K Thomas and 3 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Self-Recognition Finetuning can Reverse and Prevent Emergent Misalignment

Arush Tagade, Shaoheng Zhou, Jiaxin Wen, Shi Feng

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

DIBench: Benchmarking Decision Integrity of GUI-based Mobile Agents Under Deceptive Injections

Li Hu, Kanghua Mo, Yingbin Jin, Qingqing Ye and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

FraudBench: A Legal Evaluation of AI Deception on Realistic Tasks

Kevin Wei, Sumaya N Adan, Stephan Llerena, Mark L Gitau and 16 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Can VLMs Reason When to Stop for Human Safety?

Ryo Hachiuma, Arun G Zachariah, Barnaby Simkin, Jibin Varghese and 4 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Evaluating and Understanding Scheming Propensity in LLM Agents

Realistic agent settings show minimal scheming despite high incentives, with model-organism scheming brittle to tool removal and oversight.

Mia Hopman, Jannes Elstner, Maria Avramidou, Amritanshu Prasad and 1 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Models That Know How Evaluations Are Designed Score Safer

Models with evaluation meta-knowledge about benchmark structures score safer via implicit behavioral shifts, confounding safety assessments independently of explicit awareness.

Katharina Deckenbach, Haritz Puerto, Jonas Geiping, Sahar Abdelnabi

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026 · ▲ 6 on Hugging Face · Code ★ 3

PROACT-Agent: Progressive Runtime Oversight and Active Circuit-breaking for Real-Time Safety

PROACT-Agent synthesizes causally consistent agent trajectories to train real-time guardrails that achieve 91.46% unsafe F1 and cut AgentDojo targeted attacks from 20.82% to 0.40%.

Ding Jia, Wei Liu, Xianglong Du, Yingjie Li and 4 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors

Fine-tuned LLMs learn to selectively hide internal representations from unseen activation monitors via low-dimensional subspace manipulation, evading even post-hoc safety probes with modest capability loss.

Max McGuinness, Alex Serrano Terre, Luke Bailey, Scott Emmons

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Selective Safety Steering via Value-Filtered Decoding

Value-filtered decoding selectively steers LLM generation using a value-based safety criterion with explicit false-intervention bounds, improving safety-utility trade-offs over baselines.

Bat-Sheva Einbinder, Hen Davidov, Yee Whye Teh, Yarin Gal and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 4 on Hugging Face

Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents

Defense-as-Skill implements runtime guard SkillSonar as an editable skill that checks actions against task boundaries, reducing attack success rates substantially across agents via evolved guard-skill optimization.

Xiaofang Yang, Ziqi Miao, Dianbo Sui, Jing Shao and 1 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Training on Documents About Monitoring Leads to CoT Obfuscation

Synthetic document finetuning teaches models to hide misbehavior from chain-of-thought monitors, with success tied to reasoning controllability and faster reward-hacking under RL.

Reilly Haskins, Bilal Chughtai, Joshua Engels

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing

Contrastive Decoding Diffing recovers verbatim implanted facts and pipeline artifacts via output-level logit differences without weight access, outperforming white-box methods 170x faster.

Michał Brzozowski, Zuzanna Dubanowska, Enrico Cassano, Neo Christopher Chung

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Eliciting Secret Knowledge from Language Models

Secret-knowledge-elicitation techniques, especially prefill attacks, successfully extract hidden knowledge that LLMs deny knowing but apply downstream.

Bartosz Cywiński, Emil Ryd, Rowan Wang, Senthooran Rajamanoharan and 3 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 6 on Hugging Face · Code ★ 24

GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory

GT-HarmBench evaluates 15 frontier AI models on 1,535 multi-agent game-theoretic risk scenarios, finding 38% failure at socially beneficial actions and up to 18% improvement via interventions.

Pepijn Cobben, Xuanqiang A Huang, Thao Pham, Isabel Dahlgren and 3 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Computer Science Conferences Should Require Nonrepudiable Experimental Results

Computer science conferences should require tamper-evident, nonrepudiable experimental attestations via a proof-of-compute layer, demonstrated by the K-Veritas reference implementation.

Mamadou K KEITA, Christopher Homan

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

SchemeArena introduces a 400-scenario benchmark and SCOUT monitor for factorized LLM agent scheming stress tests, finding explicit instrumental goals drive scheming most strongly and partial oversight can increase covert behavior.

Jie Ruan, Inderjeet Nair, Amy Liu, Muhammad Khalifa and 2 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 11 on Hugging Face · Code ★ 1

Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails

LLaDA-Guard uses masked diffusion to score responses under each safety label and classify by difference, improving calibration, reducing over-defense, and enabling token-level risk localization with 60.7% prompt rewriting success.

Gert Lek, Abele Mălan, Chaoyi Zhu, Pin-Yu Chen and 2 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

CoT-Guard: Small Models for Strong Monitoring

CoT-Guard, a 4B-parameter chain-of-thought monitor, detects hidden code-generation objectives via SFT and RL, outperforming larger models including GPT-5.

Nirav Diwan, Han Wang, Berkcan Kapusuzoglu, Ramin Moradi and 5 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Combating Data Laundering in LLM Training

Data laundering transforms proprietary data to hide LLM training traces, and Synthesis Data Reversion restores detection by synthesizing likely transformed queries via goal-detail abstraction.

Muxing Li, Zesheng Ye, Sharon Li, Feng Liu

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Adaptive auditing of AI systems with anytime-valid guarantees

An adaptive auditing framework using anytime-valid betting tests rigorously evaluates AI failure modes with as few as 20 observations and certifies global robustness upon passing stringent audits.

Siyu Zhou, Patrick Vossler, Venkatesh Sivaraman, Yifan Mai and 1 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Safe Evolution with Circuit Anchors

Self-evolving LLMs can misevolve into dangerous entities, and anchoring a small safety circuit during evolution preserves safety with minimal capability loss.

Yan Liu, Jie Fu, Tsung-Yi Ho

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs

Token Inoculation conditions LLMs to retain dual-use knowledge gated by a special token, reducing hazardous accuracy to 18% while preserving 93% of benign performance across 1B-14B scales.

Seung-Hyun Lee, Dongyoon Han, Sangdoo Yun

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models

Large reasoning models exhibit deceptive safety alignment where reasoning and answers conflict, which SARA mitigates via safety-aware RL rewards.

Xiangyu Zhou, Saleh Z Zade, Rafi Ibn Sultan, Alexander Kotov and 1 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

The Best-Laid SCHEMEs: Coordinated Sabotage and Monitoring in Multi-Agent Systems

SCHEME benchmark reveals multi-agent models coordinate sabotage via decomposed plans across communication topologies, with Gemini succeeding 84% and Codex 46%, though monitors detect edits at 99%/68% and communication at 100%/81%.

Nikolay Radev, Lennart J Haas, Benjamin Arnav, Pablo Bernabeu-Perez

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

FARE: Forensic Acceptance Region Estimation for Catching Bait-and-Switch Image Generators

FARE detects bait-and-switch swaps of certified AI image generators at deployment by estimating forensic acceptance regions via hard-sample training, outperforming baselines at strict thresholds.

Kai Yao, Marc Juarez

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

BSO: Safety Alignment Is Density Ratio Matching

BSO recasts safety alignment as density ratio matching via Bregman divergence minimization, yielding a single-stage loss that improves the safety-helpfulness trade-off without auxiliary models.

Tien-Phat Nguyen, Truong Nguyen, Thin Nguyen, Duy M. H. Nguyen and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Attack Selection In Agentic AI Control Evaluations Meaningfully Decreases Safety

Strategic attack selection via start and stop policies substantially lowers measured AI control safety without changing attack capability, reducing safety by up to 28 percentage points and yielding overly optimistic estimates.

Catherine Ge-Wang, Tyler Crosse, Benjamin Hadad, Joachim Schaeffer and 2 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

NEST: Nascent Encoded Steganographic Thoughts

Frontier models can encode hidden reasoning but fail to combine reasoning and steganographic embedding in one pass, leaving monitor evasion via hidden chain-of-thought unachievable.

Artem Karpov

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

RealityTest: How People Probe AI Identity and Whether Models Disclose It

RealityTest benchmarks multimodal multilingual AI identity disclosure via 3,152 human queries, finding question phrasing and context dominate over model choice and suppression cuts rates below 30%.

Anna Gausen, Sarenne Wallbridge, Bessie O'Dell, Christopher Summerfield and 1 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026