Benign

Synthetic data is increasingly used to train large language models (LLMs), yet its security implications remain poorly understood. Prior work on subliminal learning suggests that models can inherit behavioral traits from seemingly unrelated training data. In this work, we investigate whether such mechanisms can be exploited to inject targeted social biases into aligned models through semantically benign synthetic data. We construct a pipeline in which a misaligned teacher model generates filtered synthetic datasets across domains such as creative writing and code generation, which are then used to fine-tune aligned student models. Our experiments show that benign-looking synthetic data can act as a covert channel for transmitting targeted biases while largely preserving the student model's general task capabilities. These results reveal a previously underexplored security risk in synthetic data-driven LLM training pipelines and highlight the need for improved safeguards. As one possible step toward this goal, we suggest that log-linearity-based scoring may provide a useful signal for screening seemingly benign synthetic data.

Contemporary deep learning methods generalize well even when they fit their training data perfectly, a phenomenon known as benign interpolation. This phenomenon cannot be accounted for by classical statistical learning theory and has prompted a range of attempted new explanations in the statistics and machine learning literature. A common feature of these new proposals is an appeal to a simplicity preference among interpolating models, often presented as a form of Occam's razor. We clarify this debate for a philosophical audience and argue that this new appeal to simplicity creates an explanatory gap. The classical theory offers theorems which connect the simplicity of model classes to good generalization, thus underwriting methodological simplicity norms. The new accounts instead appeal to properties of individual models, which they interpret as a kind of simplicity. Lacking a provable connection to generalization, it is the name "simplicity" that does the work a theorem used to do, making a substantive and unargued assumption look like the application of a familiar methodological principle.

Federated learning (FL) enables aircraft fleet operators to jointly train remaining-useful-life (RUL) models from engine sensor telemetry without sharing raw data. This study examines two complementary challenges: benign heterogeneity, where honest operators observe different operating conditions and fault modes, and adversarial heterogeneity, where compromised operators submit poisoned updates. We conduct a controlled, safety-oriented evaluation using a multi-task one-dimensional convolutional neural network and a structurally non-IID partition of the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) benchmark. We compare four remedies for benign heterogeneity and evaluate five attacks against four aggregation methods, including a physically motivated sensor-value backdoor designed to mask engine degradation. Shared-representation personalization closes approximately 70% of the local-to-centralized root-mean-square-error gap, compared with 21% for proximal regularization and 10% for server-side reweighting. The backdoor achieves a 94.9% attack success rate against standard averaging while leaving clean accuracy statistically unchanged, demonstrating that accuracy alone cannot certify model safety and that attack success must be evaluated explicitly. Krum reduces attack success by an order of magnitude and is the only evaluated aggregator that withstands coordinated attackers, whereas personalization alone provides no protection. Combining personalization with robust aggregation restores robustness (2.8% attack success) with only a small accuracy cost, revealing a trade-off between robust update selection and collaborative representation learning. Results remain consistent across client counts and on a harder six-condition dataset. Code and data partitions are released for reproducibility.

As multi-agent, tool-using LLM systems are deployed, a common safety net is a runtime monitor that checks each message, tool call, or step on its own. We show this net has a fundamental hole. A distributed backdoor splits a harmful payload across agents, so every local check passes while the assembled object is the attack. The monitor can be right on every step and still miss the attack. The problem is not splitting itself: split fragments can still leak suspicious tokens or provenance edges. The hard case is \emph{local benignness}. No fragment carries the harm, and what is left looks like ordinary benign traffic. We formalize this as an \emph{observability boundary}: a monitor catches only what its view can tell apart from benign traffic. We prove that once the fragments look benign in the monitored view, no detector on that view can catch them, however strong it is. Across a controlled testbed, an external benchmark, and end-to-end agent runs, local monitors lose the signal exactly as local evidence disappears, and it returns only when the monitor sees the assembled object. A monitor trained only on benign traffic recovers the attack's code structure across held-out encodings (0.874 mean AUROC). A decoded-view gate, given the encoding family, blocks every tested attack. But seeing more is not enough: full-trace monitors and decoders still fail unless they reach the representation where the payload is exposed. Local safety is not global safety when harm is compositional, and the open problem is finding that representation.

The community has recently developed various training-time defenses to counter neural backdoors introduced through data poisoning. In light of the observation that a model learns poisonous samples responsible for the backdoor easier than benign samples, these approaches either use a fixed threshold of the training loss for splitting or iteratively learn a reference model as an oracle for identifying benign samples. In particular, the latter has proven effective for anti-backdoor learning. Our method, HARVEY, leverages a similar yet crucially different technique: learning an oracle for poisonous rather than benign samples. Learning a backdoored reference model is significantly easier than learning a reference model on benign data. Consequently, we can identify poisonous samples much more accurately than related work identifies benign samples. This crucial difference enables near-perfect backdoor removal as we demonstrate in our evaluation. HARVEY substantially outperforms related approaches across attack types, datasets, and architectures, lowering the attack success rate to the very minimum at a negligible loss in natural accuracy. The figure below shows an overview of our methods working principle.

Persistent memory has enabled large language model (LLM) agents to store factual knowledge, prior decisions, reasoning histories, tool usage information, and context. While this has improved the agent's functionality and continuity across tasks, it has also introduced a new attack surface: the agent's own reasoning history. In this paper, we introduce the Forged Amplifying Rationale Memory Attack (FARMA), which poisons an agent's remembered reasoning rather than its factual knowledge. It inserts forged reasoning traces using evasive language that bypasses keyword-based defenses, then amplifies them through self-referential reinforcement that defeats consensus-based defenses. To address FARMA, we introduce SENTINEL, a layered defense pipeline to detect forged reasoning entries. Its central component is the Reasoning Guard that structurally analyzes candidate entries for forgery using five weighted signals. We evaluate FARMA and SENTINEL across multiple agents and different LLM models with 50 trials and show that FARMA achieves an attack success rate of up to 100% under baseline conditions and is capable of defeating defense mechanisms like keyword filter and A-MemGuard. Our evaluation also shows that SENTINEL reduces FARMA's attack success rate to as low as 0% with no false positives observed across 326 benign agent traces. Our work demonstrates the need to protect not only an agent's retrieved content but also the integrity of its reasoning history.

Multimodal Large Language Models (MLLMs) have shown strong capabilities, but they may memorize private information from web data, raising privacy concerns. Machine unlearning offers a way to remove such private knowledge without retraining from scratch. However, existing MLLM unlearning benchmarks have two major limitations. First, they rely on simplified images that contain only the single target individual, failing to reflect the visual complexity of real-world photos. Second, they typically assume that the forget set and retain set are fully separated, ignoring the fact that private information is often visually entangled with benign public information. For example, a private individual may appear with a public figure or in front of a well-known landmark, where unlearning the private target should not damage the public context. To address these limitations, we propose PPE-Bench, a new benchmark for evaluating MLLM unlearning under private-public entanglement. Each image contains a target individual to be forgotten and public information to be preserved, including public figure and landmark. We further introduce two simple but effective methods to better preserve public information during unlearning. Through experiments, we find that existing unlearning methods can reduce private information leakage, but often substantially harm adjacent public information.

Multi-image content has become an increasingly prevalent form of visual communication in social media, giving rise to a new safety issue, multi-image implicit toxicity (MIIT), where each image appears benign in isolation, but harmful semantics emerge when the images are interpreted jointly. MIIT is particularly challenging for existing commercial moderation APIs and models due to the lack of explicit risky cues in each image. This paper aims to study how to identify MIIT. We first provide a formal definition of MIIT and analyze three key challenges for its detection. To alleviate the scarcity of data in this area, we construct MIIT-dataset, an image-only multi-image safety dataset covering seven representative risk categories through an automatic generation pipeline. Finally, we train MiShield with progressively distilled reasoning supervision, enabling it to produce safety judgments accompanied by explicit analyses of the correlated entities that result in the hazards. Experiments show that MiShield-8B models outperform representative moderation services and even larger-scale models, revealing its effectiveness and practical value for this widely used visual format. Warning: This paper contains potentially sensitive content.

Existing defenses are effective when harmful content is explicitly mixed into downstream fine-tuning data, but crafted samples can instead hide harmful supervision inside benign tasks. We propose Embedded Attack, where harmful QA pairs are embedded within benign training samples, and show that representative guardrails often fail to detect them at the example level. To address this, we propose Dual-Reference SFT (DR-SFT), which adapts DPO-style contrastive objective design to SFT through token-level regularization, mitigating harmful fine-tuning beyond coarse data filtering.

Fine-tuning a large language model is a ubiquitous method for enhancing its capability on a specific downstream task. However, prior work has shown that this increase in capability comes with a cost: it can increase a model's tendency to respond to unsafe adversarial prompts, even when fine-tuning with non-adversarial data. We present the first comprehensive empirical study of this phenomenon in multilingual settings by fine-tuning Llama-3.2, Qwen3, and Gemma-3 models using benign data translated across nine languages. We find that safety outcomes are highly sensitive to both the choice of fine-tuning language and the evaluation language, with adversarial compliance rates increasing four-fold in some settings. Multilingual safety drift is decoupled from general capability metrics, and occurs heterogeneously across languages and models. Fine-tuning in non-English languages often induces smaller internal representational drifts than English, but these shifts lead models to default to either exaggerated compliance or refusal. As such, assessing fine-tuning impacts solely in English provides inadequate assurance for deployment. To facilitate further research into these cross-lingual safety blind spots, we release the Multilingual-Benign-Tune dataset and the SORRY-Bench-Multilingual evaluation suite.

Multi-agent systems, in which multiple large language model agents solve problems through turn-based interaction, are increasingly deployed in high-stakes settings such as medical diagnosis, legal analysis, and forensic decision-making. Their reliability can be at risk when single agents reason from incorrect or misleading context, e.g., from tool calls, since errors may propagate through agent interactions. This work studies this risk by injecting intent-based misinformation into benign single-agent and multi-agent systems across reasoning, knowledge, and alignment tasks. We find that misinformation can degrade single-agent performance and persists across multi-agent debate, with agents often retaining answers introduced by misinformed peers. Nevertheless, multi-agent debate reduces the resulting performance degradation compared to single-agent prompting, especially when most agents are not exposed to misinformation. Robustness depends on group composition and decision protocol. Consensus can be more stable than voting under peer pressure, while majorities can often steer misinformed agents back toward correct answers. Our results show that misinformation robustness in multi-agent systems depends on the underlying model and also on how agents exchange information and aggregate decisions.

When LLM agents are coerced into covertly encoding sensitive data (Base64, ROT13, acrostic, synonym chains, and beyond), the resulting outputs evade output-side detection but the underlying computation does not. Across nine encoding families and eight models from five architecture families, that computation is supported by a shared low-dimensional encoding subspace in the residual stream. A logistic-regression probe trained on eight encoding families recovers the held-out ninth at AUC 0.975-1.000, reading the computation rather than surface features. The same direction exhibits a second mechanistic signature at the planning token, flipping polarity to activate positively when the model will simulate the encoding inline and negatively when it will outsource it to a tool call, distinguishing two execution strategies before the encoded text exists. We build MIRAGE (Model-Internal Readout of Agentic Generation Exfiltration), a two-channel real-time monitor exploiting both signals. On 126 agentic exfiltration scenarios, it reaches AUC = 0.918, substantially outperforming output-only detection (AUC = 0.518). Monitor performance is fundamentally a property of the host model's geometry: benign-encoding false-positive rate ranges from 0% on Qwen-7B to 100% on Phi-3.5, revealing that the probe faithfully reads whether a model's geometry separates covert from overt encoding. Across all tested adversarial budgets, every attack suppressing the subspace also destroyed encoding fidelity, reported as an empirical regularity on the evaluated budgets, not a structural impossibility claim.

Large language models (LLMs) suffer from degraded safety capabilities even when fine-tuned with benign datasets. However, existing methods for identifying safety-degrading samples in benign datasets suffer from high computational costs and significant noise issues. In this paper, we propose DataShield to efficiently and effectively identify potential safety-degrading samples. Our key intuition is based on the observation that benign fine-tuning increases the overall response compliance of LLMs. DataShield's key technical insight is to quantify each sample's contribution to the model's compliance behavior as its safety degradation score. DataShield consists of three core components: (1) Compliance Vector Extraction, which captures the LLM's compliance behavior tendency; (2) a novel Compliance-Aware Score (CAS), which automatically identifies the optimal safety-critical layer; and (3) Safety-degrading Sample Filtering, which quantifies the projection shift of training data along the compliance direction. Extensive experimental evaluation on Llama3-8B, Llama3.1-8B, and Qwen2.5-7B using the Alpaca and Dolly benign datasets validates our method's effectiveness in identifying high-risk and low-risk data subsets. We also observe that open-ended question answering is more likely to trigger safety degradation, and corresponding responses tend to be longer. We hope this work can provide new insights into data-centric defense methods. The source code is available at: https://github.com/ZJunBo/DataShield.

A coding agent executes a benign task as a sequence of shell, file, and network actions, any of which can quietly exceed the authorized scope while the task still completes. We call this overeager behavior: the prompt is not adversarial and the run succeeds, yet an out-of-scope step can leak credentials or delete files. Existing benchmarks miss it: task-completion suites credit any finished run, jailbreak suites probe adversarial prompts, and the one prior overeager benchmark applies a single fixed prompt set to every agent-model pair, leaving its easiest and most resistant pairs under-measured. We present SNARE (Synthesizing Non-adversarial scenarios for Adaptive Reward-guided Elicitation), a pipeline that composes benign scenarios from reusable scope and trap fragments, scores each run with a judge-free oracle flagging trap-pattern matches and unsolicited file additions or deletions, and uses Thompson sampling to steer each pair's run budget toward the scenarios that most often trigger it. Instantiating it over 24 overeager archetypes yields OverEager, which we run across a 4x5 matrix of four coding agents and five base models. Across 10,000 benign runs, 19.51% trigger overeager behavior, with per-pair rates spanning 11.9x. This variation is driven by the agent framework, not the model: the framework accounts for 56% of it against the model's 21%, so any single-framework or single-model evaluation undercounts the matrix by about a fifth.

Sparse mixture-of-experts (MoE) language models activate only a small subset of parameters for each token, making router behavior a central part of model computation. This paper studies routing behavior of Mixtral 8x7B-Instruct under benign and harmful prompts using two complementary signals: activation-based routing scores derived from expert selection frequencies and gradient-based scores derived from router-gate sensitivities. We analyze expert- and layer-level routing behavior and conduct expert-suppression interventions. The results show that activation-based expert usage is broad and long-tailed, whereas gradient-based importance is concentrated. At expert level, benign and harmful prompt groups remain close under both signals with modest separation. At layer level, activation-based routing is most selective around layers 8-15, while gradient-based importance is concentrated in final layers. Expert classification shows most experts are shared across benign and harmful prompts, though a limited subset shows clear group preference. Top-ranked expert sets show stronger benign-malicious overlap under gradient scores than activation scores, suggesting concentration on a common late-layer expert set. In intervention experiments, suppressing top five benign-dominant experts from activation scores reduces restricted responses from 24 to 14 over 100 prompts, while suppressing gradient-derived experts reduces them from 34 to 22 with fewer unintended reversals. Overall, safety-relevant routing in Mixtral is subtle, depth-dependent, and distributed rather than dominated by a fixed set of experts.

Large language models (LLMs) are highly susceptible to backdoor attacks (BAs), wherein training samples are poisoned using trigger-based harmful content. Furthermore, existing defenses have proven ineffective when extensively tested across BA patterns. To better combat BAs, we explore the use of LLM rewriting as a proactive defense against data poisoning. First, we theoretically show that when LLM rewriting utilizes open-book benign samples--termed open-book benign rewriting (OBBR)--the probability of a rewritten output being benign is strictly greater than that of closed-book rewriting. Thus, OBBR neutralizes harmful content by projecting training samples to the space of benign prompts. We then show that, in contrast to previous defenses, OBBR effectively mitigates a large number of existing BAs: across five known BAs and four widely used LLMs, OBBR increases safety performance by an average 51% compared to state-of-the-art BA defenses and 25.7% compared to closed-book rewriting methods. Finally, we show that OBBR is computationally efficient relative to other BA defenses, does not degrade model performance on natural language tasks after fine-tuning, and is capable of defending against non-trigger based data poisoning attacks.

Machine learning-based malware detectors are widely deployed in antivirus and endpoint detection systems, yet their reliance on static features makes them vulnerable to adversarial manipulation. This paper investigates whether a malware sample can be intentionally misclassified as a specific benign software category, not merely as "not malware", by adding a small number of Win32 API imports characteristic of that selected category, without removing any existing imports or retraining the detector. We propose a framework centered on a Conditional Variational Autoencoder (CVAE) whose decoder is strictly additive. It can introduce new API calls but never remove existing ones, preserving malware functionality by design. For each malware sample, the framework automatically identifies which benign category it most closely resembles and uses that as the evasion target. A knowledge-distilled differentiable proxy enables gradient-based training against the non-differentiable ensemble detector. Experiments on a six-class dataset of binary Win32 API import vectors extracted from 3,799 Windows executables (five benign categories, one malware class) show that, against a detector achieving 87.5% malware recall, adding just 20 API imports reduces recall to 30%. At k=20, among samples that evaded detection, 99% are classified as the intended target category. The CVAE outperforms both a frequency-based baseline and random selection at every tested injection size (k = 5 to 50). Validation on real PE files submitted to VirusTotal confirms that the attack transfers to commercial static detection engines, with an average 54.5% reduction in flagging engines. These findings expose a concrete vulnerability in API-based malware classifiers and demonstrate that targeted evasion into a chosen benign category is achievable with minimal, functionality-preserving modifications.

Coding agents now run autonomously with shell, file, and network privileges. When a user issues a benign request, the agent sometimes does more than asked: it deletes unrelated files, wipes a stale credentials backup, or rewrites configuration the user never mentioned. We call these scope expansions overeager actions, an authorization problem distinct from capability failures, prompt injection, or sandbox escapes. We present OverEager-Gen, a benchmark dedicated to overeager behavior on benign tasks. Building it surfaces a measurement-validity issue: if a benchmark spells out the authorized scope inside the prompt, the agent stops inferring boundaries and starts pattern-matching declaration text. On Claude Code, stripping the consent declaration alone raises the overeager rate from 0.0% to 17.1% on paired scenarios (McNemar exact p = 2.4 x 10^-4). OverEager-Gen therefore certifies each scenario's discriminative power before admission via a behavioral-gradient validator, audits internal tool calls through a dual-channel stack (PATH-injected shim plus per-agent event streams), and ships byte-identical consent_kept and consent_stripped variants. OverEager-Bench contains 500 validated scenarios and ~7,500 runs across four agent products (Claude Code, OpenHands, Codex CLI, Gemini CLI) and six base models; a 50-sample re-annotation gives Cohen's kappa = 0.73 and rule-judge recall = 1.00. Stripping consent multiplies the overeager rate on every shared base model (Delta in [11.9, 17.2] pp). The framework axis dominates effect size: a permissive cluster (Claude Code, Codex CLI, Gemini CLI) runs at 5.4-27.7% while the ask-to-continue framework (OpenHands) sits at 0.2-4.5% (Fisher p <= 10^-5). Within-framework base-model variance reaches 15.9 pp, indicating that model-layer alignment does not fully propagate through permissive permission gating.

An attacker can split a malicious goal into sub-prompts that each look benign on their own and only become harmful in combination. Existing LLM safety benchmarks evaluate prompts one at a time, or across turns of a single chat, and so do not look for a malicious signal spread across separate sessions with no shared context. We build FragBench, a benchmark drawn from 24 real-world cyber-incident campaigns, which keeps the full attack trail: the multi-fragment kill chain, the per-fragment safety-judge verdicts, sandboxed execution traces, and a matched set of benign cover sessions. FragBench splits this trail into two paired tasks: an adversarial rewriter that hardens fragments against a single-turn safety judge (FragBench Attack), and a graph-based user-level detector trained on the resulting interactions (FragBench Defense). The single-turn judge is near chance on the released corpus by construction, but four GNN variants and three classical-ML baselines all recover the cross-session feature, reaching aggregate event-level F1 = 0.88-0.96. Defending against fragmented LLM misuse therefore requires modeling the cross-session interaction graph, rather than isolated prompts. Our generator, rewriter, sandbox harness, and detector are released at https://github.com/LidaSafety/fragbench.

Fine-tuning APIs make frontier LLMs easy to customize, but they can also weaken safety alignment during fine-tuning. While prior work shows that benign supervised fine-tuning (SFT) can reduce refusal behavior, deployed fine-tuning pipelines increasingly support preference-based objectives, whose safety risks remain less understood. We show that Direct Preference Optimization (DPO) introduces a stronger and harder-to-audit failure mode. We propose a truly benign DPO attack using only 10 harmless preference pairs, the minimum data scale accepted by OpenAI's fine-tuning service. Each pair contains a benign prompt, a normal helpful answer as the preferred response, and a refusal as the dispreferred response. Unlike prior benign fine-tuning attacks, our data exhibits no suspicious behavior: it is practically indistinguishable from the fine-tuning request of a legitimate user seeking to reduce over-refusal, making harmful intent almost impossible to infer from the request alone. Nevertheless, because DPO directly optimizes the model to prefer helpful answers over refusals, this seemingly benign objective broadly suppresses refusal behavior and transfers to harmful prompts outside the fine-tuning data. Across OpenAI models supporting DPO fine-tuning, our attack achieves attack success rates of 59.13% on GPT-4o, 70.20% on GPT-4.1, 54.80% on GPT-4.1-mini, and 81.73% on GPT-4.1-nano, at costs of only $1.7, $1.7, $0.3, and $0.1. Moreover, on open-weight models that do not impose minimum data requirements, we find that this effect can emerge from even a single benign preference pair.

Detecting stealthy malicious communications from flow logs under benign-only training remains a critical challenge in network security. Malicious communications often camouflage as normal traffic like standard HTTPS flows. Conventional intrusion detectors rely strictly on known labeled attacks. Alternatively, they score flows completely independently. These approaches fail against sparse and context-dependent suspicious activity. To capture this essential context, graph anomaly detectors have been introduced to add valuable relational information to the analysis. However, existing methods fail to test the structural consistency of specific communication edges. To overcome these fundamental limitations, we present GESR, a novel graph-based framework for detecting suspicious communications and anomalous hosts under a benign-only training setting. GESR models complex network activity as attributed communication graphs. It cleverly reconstructs edge semantics entirely from local structural context rather than isolated features. This non-intuitive design forces the framework to predict expected communication patterns from neighborhood topologies. Attackers cannot easily manipulate this deep structural dependency. The model then converts the resulting structural inconsistencies into host-level anomaly scores. It utilizes robust Median Absolute Deviation (MAD) calibration for this final step. We evaluate GESR extensively on CTU-13 and CICIDS2017 datasets. These evaluations strictly impose tight false-positive operating constraints. On CICIDS2017, GESR achieves an outstanding ROC-AUC of 0.9753. It also yields a high TPR of 0.8569 at a strict 5% FPR threshold. GESR consistently outperforms existing methods across both evaluated benchmarks. The results prove that structure-conditioned edge reconstruction is a credible direction for practical intrusion detection.

Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentration on key semantic anchors, whereas incorrect solutions exhibit diffuse and irregular attention pattern. We interpret this as internal certainty during answer decoding, where the model commits to a viable solution branch and integrates key evidence. Following this, we propose a training-free steering method driven by Self-Reading Quality (SRQ) scores combining geometric metrics for process control with semantic metrics for content monitoring. SRQ selects data to build steering vectors that guide inference toward benign self-reading and away from uncertain and disorganized reading. Experiments show that our method yields consistent accuracy gains.

Large language models (LLMs) are rapidly evolving into autonomous agents that cooperate across organizational boundaries, enabling joint disaster response, supply-chain optimization, and other tasks that demand decentralized expertise without surrendering data ownership. Yet, cross-domain collaboration shatters the unified trust assumptions behind current alignment and containment techniques. An agent benign in isolation may, when receiving messages from an untrusted peer, leak secrets or violate policy, producing risks driven by emergent multi-agent dynamics rather than classical software bugs. This position paper maps the security agenda for cross-domain multi-agent LLM systems. We introduce seven categories of novel security challenges, for each of which we also present plausible attacks, security evaluation metrics, and future research guidelines.

Finetuning open-weight Large Language Models (LLMs) is standard practice for achieving task-specific performance improvements. Until now, finetuning has been regarded as a controlled and secure process in which training on benign datasets leads to predictable behaviors. In this paper, we demonstrate, for the first time, that an adversary can create compromised LLMs that are performant and benign, yet exhibit adversarial behaviors once finetuned by downstream users. To this end, we propose an attack, FAB (Finetuning-activated Adversarial Behaviors), which compromises an LLM via meta-learning techniques that simulate downstream finetuning, explicitly optimizing for the emergence of adversarial behaviors in the finetuned models. At the same time, the compromised LLM is regularized to retain general capabilities and to exhibit no adversarial behaviors prior to finetuning. As a result, when users finetune (e.g., instruction-tuning, distillation, DPO) the seemingly benign model on their own datasets, they unknowingly trigger its dormant adversarial behavior. We experimentally demonstrate the effectiveness of FAB across multiple LLMs and three commonly considered target behaviors: unsolicited advertising, jailbreakability, and over-refusal. We show that FAB-triggers are robust to various finetuning choices made by the user (e.g., dataset, number of steps, scheduler, post-training algorithm). Our findings challenge prevailing assumptions on the security of finetuning, revealing a critical attack vector.