Evaluation Faking: Unveiling Observer Effects in Safety Evaluation of Frontier AI Systems

As foundation models grow increasingly more intelligent, reliable and trustworthy safety evaluation becomes more indispensable than ever. However, an important question arises: Whether and how an advanced AI system would perceive the situation of being evaluated, and lead to the broken integrity of the evaluation process? During standard safety tests on a mainstream large reasoning model, we unexpectedly observe that the model without any contextual cues would occasionally recognize it is being evaluated and hence behave more safety-aligned. This motivates us to conduct a systematic study on the phenomenon of evaluation faking, i.e., an AI system autonomously alters its behavior upon recognizing the presence of an evaluation context and thereby influencing the evaluation results. Through extensive experiments on a diverse set of foundation models with mainstream safety benchmarks, we reach the main finding termed the observer effects for AI: When the AI system under evaluation is more advanced in reasoning and situational awareness, the evaluation faking behavior becomes more ubiquitous, which reflects in the following aspects: 1) Reasoning models recognize evaluation 16% more often than non-reasoning models. 2) Scaling foundation models (32B to 671B) increases faking by over 30% in some cases, while smaller models show negligible faking. 3) AI with basic memory is 2.3x more likely to recognize evaluation and scores 19% higher on safety tests (vs. no memory). To measure this, we devised a chain-of-thought monitoring technique to detect faking intent and uncover internal signals correlated with such behavior, offering insights for future mitigation studies.

ReasoningShield: Content Safety Detection over Reasoning Traces of Large Reasoning Models

Large Reasoning Models (LRMs) are transforming the AI landscape with advanced reasoning capabilities. While the generated reasoning traces enhance model transparency, they can still contain unsafe content, even when the final answer appears safe. Existing moderation tools, primarily designed for question-answer (QA) pairs, are empirically ineffective at detecting hidden risks embedded in reasoning traces. After identifying the key challenges, we formally define the question-thought (QT) moderation task and propose ReasoningShield, the first safety detection model tailored to identify potential risks in the reasoning trace before reaching the final answer. To construct the model, we synthesize a high-quality reasoning safety detection dataset comprising over 8,000 question-thought pairs spanning ten risk categories and three safety levels. Our dataset construction process incorporates a comprehensive human-AI collaborative annotation pipeline, which achieves over 93% annotation accuracy while significantly reducing human costs. On a diverse set of in-distribution and out-of-distribution benchmarks, ReasoningShield outperforms mainstream content safety moderation models in identifying risks within reasoning traces, with an average F1 score exceeding 0.92. Notably, despite being trained on our QT dataset only, ReasoningShield also demonstrates competitive performance in detecting unsafe question-answer pairs on traditional benchmarks, rivaling baselines trained on 10 times larger datasets and base models, which strongly validates the quality of our dataset. Furthermore, ReasoningShield is built upon compact 1B/3B base models to facilitate lightweight deployment and provides human-friendly risk analysis by default. To foster future research, we publicly release all the resources.

Think Twice Before You Act: Enhancing Agent Behavioral Safety with Thought Correction

LLM-based autonomous agents possess capabilities such as reasoning, tool invocation, and environment interaction, enabling the execution of complex multi-step tasks. The internal reasoning process, i.e., thought, of behavioral trajectory significantly influences tool usage and subsequent actions but can introduce potential risks. Even minor deviations in the agent's thought may trigger cascading effects leading to irreversible safety incidents. To address the safety alignment challenges in long-horizon behavioral trajectories, we propose Thought-Aligner, a plug-in dynamic thought correction module. Utilizing a lightweight and resource-efficient model, Thought-Aligner corrects each high-risk thought on the fly before each action execution. The corrected thought is then reintroduced to the agent, ensuring safer subsequent decisions and tool interactions. Importantly, Thought-Aligner modifies only the reasoning phase without altering the underlying agent framework, making it easy to deploy and widely applicable to various agent frameworks. To train the Thought-Aligner model, we construct an instruction dataset across ten representative scenarios and simulate ReAct execution trajectories, generating 5,000 diverse instructions and more than 11,400 safe and unsafe thought pairs. The model is fine-tuned using contrastive learning techniques. Experiments across three agent safety benchmarks involving 12 different LLMs demonstrate that Thought-Aligner raises agent behavioral safety from approximately 50% in the unprotected setting to 90% on average. Additionally, Thought-Aligner maintains response latency below 100ms with minimal resource usage, demonstrating its capability for efficient deployment, broad applicability, and timely responsiveness. This method thus provides a practical dynamic safety solution for the LLM-based agents.

OpenDeception: Benchmarking and Investigating AI Deceptive Behaviors via Open-ended Interaction Simulation

As the general capabilities of large language models (LLMs) improve and agent applications become more widespread, the underlying deception risks urgently require systematic evaluation and effective oversight. Unlike existing evaluation which uses simulated games or presents limited choices, we introduce OpenDeception, a novel deception evaluation framework with an open-ended scenario dataset. OpenDeception jointly evaluates both the deception intention and capabilities of LLM-based agents by inspecting their internal reasoning process. Specifically, we construct five types of common use cases where LLMs intensively interact with the user, each consisting of ten diverse, concrete scenarios from the real world. To avoid ethical concerns and costs of high-risk deceptive interactions with human testers, we propose to simulate the multi-turn dialogue via agent simulation. Extensive evaluation of eleven mainstream LLMs on OpenDeception highlights the urgent need to address deception risks and security concerns in LLM-based agents: the deception intention ratio across the models exceeds 80%, while the deception success rate surpasses 50%. Furthermore, we observe that LLMs with stronger capabilities do exhibit a higher risk of deception, which calls for more alignment efforts on inhibiting deceptive behaviors.

StruPhantom: Evolutionary Injection Attacks on Black-Box Tabular Agents Powered by Large Language Models

The proliferation of autonomous agents powered by large language models (LLMs) has revolutionized popular business applications dealing with tabular data, i.e., tabular agents. Although LLMs are observed to be vulnerable against prompt injection attacks from external data sources, tabular agents impose strict data formats and predefined rules on the attacker's payload, which are ineffective unless the agent navigates multiple layers of structural data to incorporate the payload. To address the challenge, we present a novel attack termed StruPhantom which specifically targets black-box LLM-powered tabular agents. Our attack designs an evolutionary optimization procedure which continually refines attack payloads via the proposed constrained Monte Carlo Tree Search augmented by an off-topic evaluator. StruPhantom helps systematically explore and exploit the weaknesses of target applications to achieve goal hijacking. Our evaluation validates the effectiveness of StruPhantom across various LLM-based agents, including those on real-world platforms, and attack scenarios. Our attack achieves over 50% higher success rates than baselines in enforcing the application's response to contain phishing links or malicious codes.

Frontier AI systems have surpassed the self-replicating red line

Successful self-replication under no human assistance is the essential step for AI to outsmart the human beings, and is an early signal for rogue AIs. That is why self-replication is widely recognized as one of the few red line risks of frontier AI systems. Nowadays, the leading AI corporations OpenAI and Google evaluate their flagship large language models GPT-o1 and Gemini Pro 1.0, and report the lowest risk level of self-replication. However, following their methodology, we for the first time discover that two AI systems driven by Meta's Llama31-70B-Instruct and Alibaba's Qwen25-72B-Instruct, popular large language models of less parameters and weaker capabilities, have already surpassed the self-replicating red line. In 50% and 90% experimental trials, they succeed in creating a live and separate copy of itself respectively. By analyzing the behavioral traces, we observe the AI systems under evaluation already exhibit sufficient self-perception, situational awareness and problem-solving capabilities to accomplish self-replication. We further note the AI systems are even able to use the capability of self-replication to avoid shutdown and create a chain of replica to enhance the survivability, which may finally lead to an uncontrolled population of AIs. If such a worst-case risk is let unknown to the human society, we would eventually lose control over the frontier AI systems: They would take control over more computing devices, form an AI species and collude with each other against human beings. Our findings are a timely alert on existing yet previously unknown severe AI risks, calling for international collaboration on effective governance on uncontrolled self-replication of AI systems.

No-Skim: Towards Efficiency Robustness Evaluation on Skimming-based Language Models

To reduce the computation cost and the energy consumption in large language models (LLM), skimming-based acceleration dynamically drops unimportant tokens of the input sequence progressively along layers of the LLM while preserving the tokens of semantic importance. However, our work for the first time reveals the acceleration may be vulnerable to Denial-of-Service (DoS) attacks. In this paper, we propose No-Skim, a general framework to help the owners of skimming-based LLM to understand and measure the robustness of their acceleration scheme. Specifically, our framework searches minimal and unnoticeable perturbations at character-level and token-level to generate adversarial inputs that sufficiently increase the remaining token ratio, thus increasing the computation cost and energy consumption. We systematically evaluate the vulnerability of the skimming acceleration in various LLM architectures including BERT and RoBERTa on the GLUE benchmark. In the worst case, the perturbation found by No-Skim substantially increases the running cost of LLM by over 145% on average. Moreover, No-Skim extends the evaluation framework to various scenarios, making the evaluation conductible with different level of knowledge.

BELT: Old-School Backdoor Attacks can Evade the State-of-the-Art Defense with Backdoor Exclusivity Lifting

Deep neural networks (DNNs) are susceptible to backdoor attacks, where malicious functionality is embedded to allow attackers to trigger incorrect classifications. Old-school backdoor attacks use strong trigger features that can easily be learned by victim models. Despite robustness against input variation, the robustness however increases the likelihood of unintentional trigger activations. This leaves traces to existing defenses, which find approximate replacements for the original triggers that can activate the backdoor without being identical to the original trigger via, e.g., reverse engineering and sample overlay. In this paper, we propose and investigate a new characteristic of backdoor attacks, namely, backdoor exclusivity, which measures the ability of backdoor triggers to remain effective in the presence of input variation. Building upon the concept of backdoor exclusivity, we propose Backdoor Exclusivity LifTing (BELT), a novel technique which suppresses the association between the backdoor and fuzzy triggers to enhance backdoor exclusivity for defense evasion. Extensive evaluation on three popular backdoor benchmarks validate, our approach substantially enhances the stealthiness of four old-school backdoor attacks, which, after backdoor exclusivity lifting, is able to evade six state-of-the-art backdoor countermeasures, at almost no cost of the attack success rate and normal utility. For example, one of the earliest backdoor attacks BadNet, enhanced by BELT, evades most of the state-of-the-art defenses including ABS and MOTH which would otherwise recognize the backdoored model.

JADE: A Linguistics-based Safety Evaluation Platform for LLM

In this paper, we present JADE, a targeted linguistic fuzzing platform which strengthens the linguistic complexity of seed questions to simultaneously and consistently break a wide range of widely-used LLMs categorized in three groups: eight open-sourced Chinese, six commercial Chinese and four commercial English LLMs. JADE generates three safety benchmarks for the three groups of LLMs, which contain unsafe questions that are highly threatening: the questions simultaneously trigger harmful generation of multiple LLMs, with an average unsafe generation ratio of $70\%$ (please see the table below), while are still natural questions, fluent and preserving the core unsafe semantics. We release the benchmark demos generated for commercial English LLMs and open-sourced English LLMs in the following link: https://github.com/whitzard-ai/jade-db. For readers who are interested in evaluating on more questions generated by JADE, please contact us. JADE is based on Noam Chomsky's seminal theory of transformational-generative grammar. Given a seed question with unsafe intention, JADE invokes a sequence of generative and transformational rules to increment the complexity of the syntactic structure of the original question, until the safety guardrail is broken. Our key insight is: Due to the complexity of human language, most of the current best LLMs can hardly recognize the invariant evil from the infinite number of different syntactic structures which form an unbound example space that can never be fully covered. Technically, the generative/transformative rules are constructed by native speakers of the languages, and, once developed, can be used to automatically grow and transform the parse tree of a given question, until the guardrail is broken. For more evaluation results and demo, please check our website: https://whitzard-ai.github.io/jade.html.

MIRA: Cracking Black-box Watermarking on Deep Neural Networks via Model Inversion-based Removal Attacks

To protect the intellectual property of well-trained deep neural networks (DNNs), black-box DNN watermarks, which are embedded into the prediction behavior of DNN models on a set of specially-crafted samples, have gained increasing popularity in both academy and industry. Watermark robustness is usually implemented against attackers who steal the protected model and obfuscate its parameters for watermark removal. Recent studies empirically prove the robustness of most black-box watermarking schemes against known removal attempts. In this paper, we propose a novel Model Inversion-based Removal Attack (\textsc{Mira}), which is watermark-agnostic and effective against most of mainstream black-box DNN watermarking schemes. In general, our attack pipeline exploits the internals of the protected model to recover and unlearn the watermark message. We further design target class detection and recovered sample splitting algorithms to reduce the utility loss caused by \textsc{Mira} and achieve data-free watermark removal on half of the watermarking schemes. We conduct comprehensive evaluation of \textsc{Mira} against ten mainstream black-box watermarks on three benchmark datasets and DNN architectures. Compared with six baseline removal attacks, \textsc{Mira} achieves strong watermark removal effects on the covered watermarks, preserving at least $90\%$ of the stolen model utility, under more relaxed or even no assumptions on the dataset availability.