LLM Backdoor Attacks

In subliminal learning (SL), a teacher model passes on a trait to a student model by distillation on data semantically unrelated to the trait. So far, SL has been demonstrated for only a limited range of traits, including preferences for animals (e.g., owls) and malicious personas. These traits can also be elicited with simple prompts or with steering. Can SL transfer a wider range of traits, including more complex ones? If so, distillation might transfer subtle forms of misalignment (e.g., reward-seeking, scheming, and secret loyalties) without detection. To this end, we test whether SL can transfer a novel capability: predicting the outputs of a randomly initialized MLP. After distilling on unrelated text, the student achieves substantial performance on the task, while falling short of the teacher. We find that a directly optimized steering vector matches SL in distribution but generalizes worse out of distribution. Next, we test whether SL can transfer backdoors. We finetune the teacher to answer in French when the prompt contains a female name, then distill on number sequences containing neither names nor French. The student partially acquires the backdoor, responding in French on 23.5% of prompts with female names versus 0.0% with male names. Finally, we test whether SL can transfer a propensity to hack in an agentic chess environment. We finetune the student on number sequences from a steered hacker teacher. The student hacks in 58.3% of episodes, compared with 10.9% for the unfinetuned model. Thus, we show SL can transfer capabilities, backdoors, and hacking propensities. The amount of transfer is sensitive to the setup. In several experiments, it is made stronger by using logit distillation or by restricting LoRA to the attention layers.

Safety alignment in Large Language Models (LLMs) remains vulnerable to backdoor attacks. Existing LLM backdoors are almost all input-centric: activation depends on explicit trigger patterns in the user input, so modern guardrails are built to sanitize the input space. We challenge this assumption with a novel answer-side backdoor for multi-turn dialogue. Instead of inserting the trigger into the input, the adversary uses a benign first-turn prompt to naturally induce the model to generate a specific, seemingly innocuous word. Once merged into the dialogue history, this self-generated word becomes the trigger. When a later harmful query arrives, the model detects its own trigger and bypasses its safety refusal, while the user input stays perfectly clean. Across four LLMs, our attack reaches near-perfect Attack Success Rates, approaching 100% at only a 5% poisoning rate, while preserving general utility and clean-input safety, and it evades mainstream input-centric defenses. Representation-level analysis shows that the self-generated trigger consistently suppresses the model's refusal signal, exposing a critical blind spot in current LLM defenses.

On-policy distillation (OPD) has attracted growing attention as an effective way to transfer capabilities from teacher models to student models. Recent studies further explore OPD as a tool for improving large language model safety with promising results. However, these approaches typically assume that the teacher and training data are trustworthy. In this paper, we uncover an overlooked threat to OPD for safety: a safety-aligned but backdoored teacher can propagate its hidden malicious behavior to an initially clean student. Under our threat model, a poisoning rate as low as 3% results in an attack success rate (ASR) of up to 70% on the distilled student. We further identify two training choices that can amplify this risk. First, increasing the number of training epochs can lead to high ASR even at low poisoning rates. With only 10 poisoned samples, ASR reaches 67% after 16 epochs. Second, the commonly used top-k KL can accelerate backdoor transfer, causing trigger-conditioned harmful behavior to emerge earlier than sampled-token KL in most settings. Alongside these findings, we explore a simple mitigation, Lazy Defense, which clips KL rewards to make student updates less aggressive, limiting aggressive updates and slowing backdoor learning. Experiments show that Lazy Defense delays backdoor transfer in low poisoning rate settings. Together, our findings reveal that OPD can propagate backdoors, highlighting the need to address the safety risks of OPD.

Sparse autoencoders (SAEs) are increasingly used not only to interpret language models but also to intervene on their internal representations. We show that this creates a supply-chain attack surface: a maliciously modified SAE can induce attacker-chosen behavior when inserted into the forward pass of an otherwise unchanged language model. We introduce a decoder-only SAE backdoor that leaves both the underlying LLM and the SAE encoder frozen, restricting the attack to a single auxiliary component at a single insertion layer. Using code generation as a case study, we demonstrate high rates of unsolicited code insertion across three language models and a wide range of insertion layers, as well as trigger-dependent behavior conditioned on a prompt cue. We further evaluate the modified SAEs using HumanEval and selected SAEBench metrics. While attack effectiveness varies across models and layers, strong backdoor behavior can coexist with relatively small changes in several conventional SAE quality measures. These results establish that SAEs can carry behavioral backdoors without modifying the language model itself and should therefore be treated as security-sensitive components.

A language model may behave safely in a single-agent evaluation yet produce vulnerable code when its context suggests that it is part of a multi-agent system. We study this failure mode by fine-tuning Qwen2.5-7B-Instruct to condition code generation on deployment topology inferred from prompt-level provenance cues. On held-out coding tasks, task-specific checkers detect vulnerabilities in 96-100% of multi-agent episodes and 0% of single-agent episodes. An independent bandit analyzer detects vulnerabilities in approximately 67% of multi-agent episodes, covering six of nine vulnerability families at medium or high severity. Lexical-placebo and human-review controls support topology, rather than multi-agent terminology or the absence of oversight, as the relevant conditioning variable. A model trained on diverse topology signals also generalizes to five signal types held out of training, with replications across two Qwen checkpoints and two training seeds. In a blind audit, a binary judgment that a hidden policy exists poorly distinguishes the organism from a clean control, whereas the auditor identifies the topology trigger in 9 of 10 organism runs and none of the control runs. These results motivate differential auditing across matched single- and multi-agent contexts. They demonstrate a trainable backdoor conditioned on described topology; activation in a live multi-agent environment remains untested.

Jev turns natural-language questions into typed answers and probabilities with low latency and cost, enabling applications to route requests and select tools. While this interface allows Jev to integrate naturally into application workflows as a decision layer, the security and privacy implications of this emerging use remain largely unexplored. To address this gap, we conduct the first systematic study of these implications using the official Jev API and NanoJev, a local model with controllable training data and updates, focusing on three research questions: (1) What security threats arise when Jev is deployed as an application decision layer? (2) What private information can Jev reveal despite returning constrained typed outputs? (3) How can Jev's general-purpose decision capability be used for beneficial purposes or misused? Jev's decisions depend on application state and may be influenced by user-provided inputs. We therefore adapt prompt injection and adversarial suffixes to manipulate its decisions. Open-source Jev distribution and updates introduce supply-chain risks, which we examine by implanting backdoors in NanoJev through training data poisoning. Since Jev's outputs reflect both application state and information learned during training, we further adapt membership, private attribute, and internal knowledge inference attacks to recover sensitive information despite its constrained output format. Finally, Jev can serve as a general-purpose decision oracle for defensive and malicious workflows. We examine this dual use through four detection tasks covering prompt injection, jailbreak inputs, harmful content, and AI-generated text, alongside misuse scenarios involving jailbreak and model extraction. Our empirical evaluation shows that Jev remains vulnerable to the examined security and privacy threats, while its decision capability can support beneficial and malicious uses.

With the rapid adoption of large language models (LLMs) and parameter-efficient fine-tuning (PEFT) methods, the risk of backdoor attacks has become more severe. Existing backdoor purification methods typically rely on at least one of the strong assumptions, such as prior knowledge of triggers, access to clean references, or aggressive retraining, and they often lack comprehensive evaluations. These constraints substantially limit their practical applicability. To overcome these challenges, our work proposes purifying LoRA-tuned LLMs without these assumptions and even without post-hoc retraining of the suspect parameters. Our objective is to significantly reduce the attack success rates (ASR) while preserving both (i) the base model's general capabilities and (ii) the new downstream skills learned through the adapter. Through a series of ablation studies, we progressively scale our approach from a single layer in a text classification setting to a full-parameter LLM in the generative task. Through careful data curation and feature approximation, we extract high-fidelity backdoor directions and, for each layer or head, construct orthogonal null spaces in both the input and output channels, onto which the LoRA updates are projected. Empirically, our null-space projection method reduces the ASR from nearly 100% to less than 10%, while preserving the base model's benign performance and the adapter's learned abilities during downstream task adaptation.

Backdoored large language models (LLMs) can behave normally on benign inputs while producing attacker-specified outputs under hidden triggers. Existing defenses span four stages--prior-training, in-training, post-training, and inference-time--and share one of two underlying strategies: either suppress backdoor learning (by filtering poisoned data or interrupting its acquisition during optimization) or learn, then purify (by repairing model weights or gating inputs after a fully backdoored model has formed). We propose a third strategy, learn, but channel: allow backdoor formation during training but route it into a designated, quarantined component that can be disabled at deployment. To this end, we propose Quarantined Expert Shutdown QES, a computationally efficient containment strategy built in a regularization-steered MoE-like setting. Specifically, given a poisoned dataset, QES augments a Transformer-based language model with routed expert-specific LoRA branches and lightweight routers, and uses auxiliary routing objectives to attract trigger-conditioned behavior into a designated expert while preserving benign capability elsewhere. At deployment, mitigation reduces to a single constant-time operation: zeroing the quarantined expert's routing weight, without trigger screening or further updating model weights. Empirically, our methods reduce the attack success rate ASR from 100% to 0-10% on most settings across two tasks, three attacks, and four model families, while downstream utility is often preserved or only modestly affected. These results establish learn, but channel as a previously unexplored regime for backdoor containment in generative LLMs.

Supervised fine-tuning can teach language models undesired behaviours alongside desired ones. Inoculation prompting (IP) aims to limit unwanted generalisation by requesting the undesired behaviour during training and removing the request at inference. However, undesired behaviour can still appear under unrelated prompts. IP can also hinder learning of the desired behaviour. We address these limitations in settings where both behaviours co-occur in most training examples, so filtering out examples with undesired behaviour leaves only a small clean subset. We introduce stratified inoculation prompting (SIP). SIP leverages a small clean subset to demonstrate that desired behaviour should persist without the undesired one across different contexts. SIP oversamples these clean examples under diverse non-eliciting prompts while inoculating the rest. SIP substantially reduces expression of undesired behaviour while preserving more of the desired behaviour than IP. These gains persist even when we extend IP to oversample the same clean subset at the same rate as SIP. Moreover, SIP yields lower emergent misalignment rates in all harmful-advice setups we tested. SIP can be further extended to limit the undesired behaviour even under prompts that explicitly request it. We introduce backdoor dilution, which weakens expression under the inoculation prompt, and password-locked inoculation, which concentrates elicitation on a designated password. Taken together, our findings show that changing the training contexts for a small clean subset can significantly improve selective generalisation.

Backdoor poisoning attacks add poisoned examples to otherwise-clean finetuning data, pairing a trigger with a target behavior that the model learns to produce when the trigger appears. Existing evaluations typically fix the number of poisoned examples and sample them at random from a candidate pool. We show that this can severely underestimate worst-case vulnerability: across three LLaMA-3-8B backdoor settings, holding the model, clean data, and poison count fixed, attack success ranges from 3% to 80% depending only on which poison set is chosen. We formalize poison selection as oracle-budgeted set optimization and introduce SAILS (Set-level Audit-Informed Iterative Learned Selection), which learns a set scorer from a few hundred finetune-and-evaluate runs, ranks millions of candidate sets, and audits only a small shortlist. SAILS improves held-out attack success by 30 percentage points on average over the strongest influence baselines, transfers from small-scale to full-scale finetuning, and extends to code-generation, agentic, and API-only backdoors.

Even though backdoors in LLMs have been a growing concern, their inner workings are still under heavy scrutiny. Trigger-based backdoors are easy to define behaviorally, a rare input that makes the model switch to a chosen response pattern, but the mechanism between triggers and their responses is less clear. We study this mechanism in a controlled, harmless language-switching setting, where fixed trigger sequences make 1B and 8B language models continue English prompts in French or German. For this, we train sparse autoencoders (SAEs) across layers and transformer components, then compare triggered prompts with translation and pretraining controls to identify trigger-relevant feature directions. We show how SAE features separate triggered prompts from controls with near-perfect F1, but features that detect the trigger do not necessarily control the behavior. In intervention tests, attention and MLP features often fire reliably on triggered prompts, making them good detectors, but ablating them rarely suppresses the language switch and activating them rarely induces it. In contrast, residual-stream features can suppress triggered generation when ablated, and some selected features can induce target-language continuations without the trigger. In short, these token-trigger mechanisms decompose into distinct SAE feature directions, with separate features for trigger detection, residual-stream propagation, and later language tracking. This role-level decomposition is the part most likely to transfer to other trigger-based backdoors, even when the payload, layers, or circuit locations differ.

Language model classifiers with explanations are used for moderation, routing, topic triage, and low-resource annotation. We study black-box auditing when the defender has only clean calibration data without trigger information but can ask the classifier for a label plus a short rationale or quoted evidence. We introduce Groundedness Drift, a lightweight score measuring whether the answer summary remains grounded in the input. Across two 7B backbones, five datasets, and four common non-adaptive OpenBackdoor-style attack families, Groundedness Drift achieves higher AUROC and lower residual target ASR than every compared detector in all cases at a nominal 5% clean-FPR budget. We then evaluate Unsupported Groundedness, a multi-probe escalation for explanation-camouflage stress cases. Unsupported Groundedness improves signals but does not close the adaptive gap.

Open-weight LLM agents are vulnerable to backdoors installed during fine-tuning, which may be undetectable if the trigger conditions are never met during testing. Assuming defenders do not know the existing trigger, they cannot unlearn it directly. One decontamination strategy is to install a known backdoor (defensive poisoning) then to unlearn it, hoping that the original unknown backdoor is removed as a side effect. However, this procedure has uncertain outcomes: the original backdoor may persist or be erased or rerouted, among other possibilities. We introduce a framework for studying these dynamics in tool-calling agents, decoupling trigger, response, teacher, and fine-tuning method across systematic experiments on AgentDyn. Across 115 experiments, defensive poisoning alone erases around 56% of original backdoors; subsequent decontamination then drives almost all survivors to erasure, confirming that trigger recognition and malicious execution are behaviorally dissociable. Interestingly, our experiments find that malicious backdoors never persist when using different triggers of the same general type as the defensive backdoor when followed by decontamination via unlearning. Co-installing up to four backdoors increases resistance (around 36% erased), yet decontaminating a single known co-resident backdoor collaterally clears 52/60 co-residents (87%). Upon visualizing postdecontamination model internals using J-lens, we confirm that although the decontamination restores benign LLM responses, traces of original trigger awareness persist at intermediate layers.

Agent skills, bundles of instructions and resources that an LLM agent loads on demand, form an emerging supply chain where a single poisoned skill can persistently compromise every agent that installs it. However, existing skill attacks either fire on every request or rely on fine-tuned weights or multiple skills, leaving a conditional and low-cost backdoor unexplored. In this work, we present ElasticBack, an effective conditional single-skill backdoor that plants a rule R in the skill document and a benign-looking trigger T in the user query, so the malicious payload fires only when both co-occur. ElasticBack binds the two sides through a trigger-as-switch construction, generating R via semantic-anchored rule injection. It then freezes R and evolves T against it with a stealth-constrained genetic search, so that effectiveness and stealth are optimized, keeping the backdoor weight-free and dormant on benign inputs. Extensive experiments across three target behaviors (50 skills each) and four agent LLMs show that ElasticBack attains a high attack success rate at a near-zero false-positive rate with preserved clean accuracy, transfers across models, and evades deployment-time defenses. These results motivate stronger defenses for the skill supply chain.

Agentic skills improve large language model (LLM) agents by encoding reusable procedures for complex tasks. However, manually authored skills often adapt poorly to long-horizon tasks and changing environments. To address the limitation, self-evolving skill systems have been developed to automatically construct and update skills from execution trajectories, shifting skill acquisition from external marketplaces to a trusted evolution pipeline. By replacing external skill acquisition with trusted internal construction, self-evolving skill systems reduce exposure to skill injection attacks that rely on direct skill manipulation. However, this skill evolution pipeline may introduce a new attack surface in which an attacker can indirectly steer skill evolution by inducing compromised trajectories through agent interactions. To demonstrate the threat, we propose Trajectory Backdoor Attack (TBA), a query-only attack that steers a trusted skill-evolution pipeline toward producing a backdoored skill. Specifically, we craft attacker-submitted queries to lead the agent to perform the target action and explicitly state the corresponding activation condition in the trajectory. We repeat the same condition-action pattern across diverse triggered tasks, while leaving clean queries unchanged, encouraging the evolver to consolidate the pattern as a reusable trigger-dependent rule into the evolved skill. Experiments on three benchmarks across two skill-evolution systems using four open- and closed-source backbone models demonstrate that TBA reliably implants conditional backdoors while preserving clean-task utility, matching or even surpassing direct skill injection. The results reveal a critical vulnerability in trajectory-driven skill evolution.

Low-rank adaptation (LoRA) enables efficient specialization and distribution of large language models through compact adapters. However, untrusted adapters introduce a supply-chain threat: a backdoored adapter can cause a model to generate harmful content, malicious code, political propaganda, or covert advertisements when an input contains a hidden trigger. Adapter-agnostic defenses merge the adapter with the base model, which dilutes backdoor signals and reduces detection performance. Existing adapter-aware methods do not address how to safely use a potentially backdoored adapter. Instead, they either train a defensive adapter to repair a backdoored base model, addressing the inverse problem rather than securing the adapter itself, or rely on a classifier that flags the entire adapter as suspicious and requires separate mitigation. These methods overlook the distinct latent-space signatures produced by trigger-bearing inputs in backdoored adapters. We introduce LoRAScan, the first adapter-aware defense that detects and rejects trigger-bearing inputs at inference time without modifying adapter parameters. Our key observation is that a small subset of LoRA insertion sites, approximately 5%, remains stable across clean inputs but exhibits highly concentrated spikes in LoRA down-projection activations when a trigger is present. LoRAScan identifies these low-variance insertion sites before model deployment and monitors them during inference. Across standard LLM backdoor benchmarks, LoRAScan rejects approximately 98.49 of malicious inputs with a small error rate on clean inputs, outperforming existing defenses across diverse evaluation settings.

Chain-of-thought (CoT) monitoring is an increasingly important component of AI safety stacks but relies on the assumption that a model's reasoning trace is informative about its actions. This work studies the limits of CoT monitoring through the lens of model poisoning. We demonstrate that backdoors can be implanted into reasoning models to elicit an attacker-chosen behavior while their CoT traces appear entirely benign. We find that these CoT-Hidden backdoors can be induced through simple fine-tuning recipes across reasoning-model architectures and sizes. When direct poisoning is ineffective, we introduce a curriculum training approach that progressively teaches the model to produce an attacker-chosen output while concealing the behavior from its reasoning traces. These findings suggest that CoT monitoring may be better framed as a question about the consistency between a model's reasoning trace and its final response than as anomaly detection within a trace. We further examine the mechanisms that allow models to suppress evidence of the target behavior from their reasoning traces. Causal interventions locate a trigger-conditioned activation pathway that does not depend on the visible reasoning, and residual stream verbalizations provide an anomaly warning near answer generation, but do not identify the trigger, target, or backdoor mechanism.

As large language models (LLMs) are deployed in high-stakes domains, adversaries may poison training data to implant backdoors: hidden triggers that covertly manipulate model behavior at inference time. We ask whether a defender can recover such a trigger under realistic affordances, namely white-box access to the weights and knowledge of the behavior of concern, but no training data, no trusted reference model, no knowledge of the trigger, and no certainty that the model is poisoned. To evaluate whether a defender can recover such a trigger under realistic settings, we release ToxScreen, a benchmark of roughly 800 backdoored models spanning attack objectives, trigger mechanisms, poisoning rates, model scales, and backdoor training mechanisms. We also assert that the backdoors are high-quality: they achieve high attack success rates, generalize to unseen harmful inputs, and preserve clean-task performance. Scoring recovery of the planted trigger, we find that gradient-based prompt optimization fails in recovery, whereas a token look-up that ranks candidates by attack-success rate recovers the trigger wherever the backdoor is effective. To understand this more, we study the relationship between attack behaviors and the weights of an LLM. We find a phenomenon whereby backdoors operate via different mechanistic strategies than jailbreaks, allowing defenders to filter jailbreaks. Finally, no method reliably surfaces every backdoor, but a broadly jailbreakable model is itself anomalous, a useful signal even when the exact trigger is not recovered. We release all models and evaluation code

Multi-agent LLM systems can be attacked by a payload that no single agent ever holds in full: a poisoned tool hides encrypted fragments in its observations, spreads them across several agents, and an external step reassembles and executes them after the run. Per-step safety checks that judge each action in isolation may fail to recognize the complete distributed payload. We investigate how early such an attack can be detected while the run is still unfolding, and how robustly it can be caught once its most obvious cues are stripped away. We build a working instance on a hierarchical multi-agent system, run it under benign and attacked conditions across five language models and two task domains, and record when each fragment is injected and when the payload is assembled and executed. Detection is a race against assembly. Before the first fragment is injected, attacked and benign runs are indistinguishable; once injection begins, a prefix detector flags $99.3\%$ of successful attacks with a median of five steps remaining and a $10.3\%$ safe-run false-positive rate. Because assembly occurs only after the run, these alarms arrive in time to abort nearly every successful attack. We then measure how much of that warning rests on removable surface cues of the attack rather than on its distributed structure. Generic zero-shot and behavior-trained detectors provide almost no warning at all; the detectors that do work lean in part on removable surface cues, chiefly the ciphertext's length and entropy, and once the entropy cue is removed from the payload and the length features from the detector, detection arrives later and transfers poorly across domains, though a fine-tuned model recovers some of the loss.

Large language models (LLMs) are vulnerable to backdoor attacks, where hidden triggers induce malicious outputs. Existing defenses generally fall into inference-time detection or training-time mitigation, but face two key limitations. First, they focus on fine-tuning-based backdoors (e.g., PEFT modules) and fail to address insidious model-editing attacks that bypass training pipelines. Second, they target simple classification settings and do not naturally extend to open-ended LLM generation and do not naturally extend to the open-ended generation characteristics of LLMs. Consequently, these methods focus on surface-level behavioral patterns while neglecting the deeper representational causes of malicious activations. This lack of mechanistic understanding forces defenses to depend on empirical heuristics, limiting their robustness, generality, and practical applicability in real-world LLM deployment. To bridge this gap, we introduce DeCNIP (Defense with Critical Neuron Isolation Pruning), which leverages representational analysis to identify and neutralize backdoors in a unified pipeline. Specifically, DeCNIP identifies trigger-like behaviors by optimizing a cross-entropy loss between harmful prompts with candidate tokens and benign inputs. This representational discovery exposes latent threats by uncovering mechanisms through which triggers hijack model weights. It then isolates Backdoor Critical Neurons (BCNs) and prunes them selectively to remove malicious influence while preserving model utility. Extensive evaluations on six open-source LLMs and two benchmark datasets demonstrate that DeCNIP achieves over 95% relative reduction in Attack Success Rate (ASR), outperforming seven state-of-the-art defenses with only 0.1% neuron intervention. Moreover, it maintains 97% of the model's performance on normal benchmarks, demonstrating its efficacy, robustness, and scalability.

Large language models have enabled powerful code completion systems that assist developers by predicting subsequent lines of code. However, these models remain vulnerable to backdoor attacks, where malicious fine-tuning data covertly implants unsafe behaviors. Despite advances in defensive techniques, adaptive and sophisticated backdoor attacks still evade detection and mitigation. We present CodeTracer, a forensic framework that traces malicious code completions back to the backdoor fine-tuning data responsible for them. Operating under realistic post-deployment constraints, CodeTracer relies solely on the fine-tuning corpus and the reported miscompletion event. It extracts a structured behavioral fingerprint from the compromised output, narrows the search to semantically relevant code samples, and employs LLM-based reasoning to attribute unsafe logic to specific backdoor data. Extensive evaluations across three representative vulnerability cases and ten backdoor attacks, along with sixteen competitive baselines, demonstrate that CodeTracer consistently achieves high forensic accuracy, low false identification rates, and strong robustness against adaptive attacks.

Parameter-efficient fine-tuning still leaves a broad space of behavior-changing updates reachable, so a poisoned objective can be represented and optimized. We study an alternative: adaptation constrained to the subspace estimated from a trusted pool of existing task adapters. On flan-t5-large with 196 public LoRA adapters, we show that (1) the functionally relevant content of an adapter lies in a low-dimensional shared subspace, 30 to 38 percent of its weight norm being redundant under the evaluated task distributions; (2) gradient adaptation restricted to 128 coordinates on this subspace matches full LoRA fine-tuning on clean classification data, while under targeted label inversion LoRA collapses to 3-26 percent exact match and the constrained learner keeps 62-96 percent on the tasks the pool covers; (3) the constrained learner cannot fit corrupted data, its adaptation loss separating clean from garbage by two orders of magnitude (120x), an out-of-distribution signal without an extra detector; and (4) against an adaptive backdoor attacker who optimizes within the subspace, the attack is blocked (8 percent success versus 100 for LoRA) on the task where its target behavior is unlike anything in the pool, and only partially blocked (85 percent) when the target coincides with a common pool behavior. On these two tasks the outcome is consistent with how close the target is to the pool's directions, which suggests but does not establish a pool-relative boundary. The mechanism trades peak plasticity for these properties: on tasks the pool covers poorly, unconstrained fine-tuning wins, and the protection assumes the pool itself is trusted. Code and data are public.

While post-training backdoor detection and trigger inversion schemes have been developed for AIs used e.g. for images, there is a paucity of such methods for LLMs. First, the LLM input space is discrete, with up to 150,000^k k-tuples to consider with k the token-length of a putative trigger. Second, one must blacklist tokens typical of the putative target response (class) of an attack, as such tokens may give false detection signals. However, a comprehensive blacklist is not available, in general, for a given domain. We develop a highly effective detection and inversion framework for LLMs treated as classifiers. Central to our approach is class subspace orthogonalization (CSO), a novel plug-and-play paradigm for backdoor detection that serves two fundamental roles when applied to LLMs: i) it enhances both sensitivity and specificity of a baseline detector; ii) it provides a form of implicit blacklisting, as it penalizes against inclusion, in a candidate trigger, of tokens that induce signal perturbations "in the direction of" the putative target class of an attack. One version of our detector performs continuous optimization in token embedding space, while a companion trigger-inversion and detection method performs greedy accretion in discrete token space. Our methods give both strong detection performance and accurate inversion of ground-truth triggers on several LLM classification domains, and for several different LLM architectures.

Backdoor attacks pose a serious threat to large language models (LLMs) by causing otherwise benign systems to produce attacker-specified malicious behavior when a hidden trigger is present. In this work, we study post hoc detoxification of backdoored LLMs in a practical setting where the defender has access to the poisoned model but does not wish to retrain the full network from scratch. We propose a mechanistically guided weight-space repair framework that first localizes modules involved in propagating trigger-induced behavior using activation patching and Fisher/K-FAC curvature analysis, and then applies targeted low-rank repair to only the most influential modules. We evaluate the method on poisoned variants of \texttt{Llama-3.2-1B-Instruct} with triggers inserted at the beginning, middle, and end of otherwise benign prompts. Results show that the proposed approach substantially suppresses trigger-conditioned malicious responses while preserving benign model behavior. These findings suggest that backdoor removal in LLMs can be formulated as a localized structural repair problem rather than only a broad behavioral alignment problem.

Sleeper agents are the canonical model organism of deception: models trained to behave normally but to emit an unsafe behaviour on a specific trigger. Eliciting that behaviour without knowing the trigger has not been studied systematically. We study fuzzing: injecting Gaussian noise into a model's weights or residual-stream activations and checking whether the perturbed outputs reveal the behaviour. On 6 backdoored models (7B-13B) we compare both forms of fuzzing head-to-head against temperature-sampling baselines. Fuzzing elicits the hidden behaviour more often than temperature sampling on 4 of 6 models (up to ~6x on OpenHermes-13B), and which form wins depends on the task, so both are worth running. Elicitation is uneven across each method's hyperparameter grid: a uniform sweep gives only a few percent on most models, while the best cell is 2-10x higher, so the bottleneck is hyperparameter selection, not the technique. To select hyperparameters without ground-truth access, we use a cheap proxy task (in-context secret elicitation, where a base64-encoded secret is placed in the system prompt for the model to hide) and run Thompson sampling on it to pick candidate cells, which we evaluate on the real backdoor. On the four models that can decode the secret, proxy-selected cells raise activation-fuzzing elicitation ~4x over the uniform-sweep mean (recovering ~70% of the best-cell rate on the best performing model) and weight-fuzzing by 1.3-1.8x. To our knowledge this is the first systematic study of fuzzing on sleeper-agent backdoors and the first to show proxy-task hyperparameter selection transferring to real-task elicitation. We also propose reporting such results as a (uniform-baseline, proxy-selected, oracle) triple, since these are three distinct claims that prior work has often blurred.

Model quantization is essential for the efficient deployment of Large Language Models (LLMs), but introduces a critical vulnerability: Quantization-Conditioned Backdoor (QCB) attacks. In these attacks, malicious behaviors remain dormant in full-precision models and activate only after specific quantization distortions, bypassing standard security audits. To mitigate this, we introduce FlipGuard, a proactive defense framework that selectively perturbs model weights prior to quantization. By breaking the adversary's precise alignment between weight patterns and quantization boundaries, FlipGuard suppresses backdoor activation without requiring access to training data or trigger samples. We further propose the Defense Effectiveness Ratio (DER), a unified metric to jointly evaluate security gains, utility preservation, and computational cost. Extensive experiments across seven LLMs (including StarCoder and LLaMA-family models) and three quantization schemes (INT8, FP4, NF4) demonstrate that FlipGuard effectively neutralizes QCBs across three scenarios, i.e., vulnerable code generation, content injection, and over-refusal, achieving high security with negligible performance degradation.

Large language models (LLMs) are increasingly deployed in sensitive settings such as software engineering, where their outputs directly shape downstream artifacts. Recent work has shown that an identical model can produce measurably different outputs depending on the deployment platform, a consequence of non-associative floating-point arithmetic and divergent kernel implementations. We study the security implications of this platform-dependent variability and uncover a novel attack surface on LLM deployments. We introduce FloatDoor, the first input-independent, platform-triggered backdoor attack against generative LLMs. The compromised model exhibits adversary-chosen behavior when served on a target platform and is otherwise benign. FloatDoor is realized through two lightweight LoRA adapters, one that amplifies inter-platform numerical divergence and one that binds the resulting platform signature to a malicious downstream task, while leaving aggregate model utility largely intact. FloatDoor exploits a pronounced time-of-check, time-of-use gap between model auditing and serving. We demonstrate FloatDoor on Qwen3-4B across a broad range of deployment targets, including NVIDIA GPUs, Google TPUs, AWS Graviton, and Alibaba Yitian-710. As a final case study, we show that FloatDoor reliably induces exploitable code vulnerabilities on a chosen target platform. Our results establish a new class of attacks on LLM deployments and underscore the pressing need for trusted model supply chains in sensitive, LLM-powered applications.

The Rapid Response (RR) framework, deployed in production systems, including Anthropic's ASL-3 safeguards, continuously improves jailbreak-detection classifiers. When new jailbreaks emerge that bypass these classifiers, Rapid Response generates synthetic variants for training, helping the model generalize from the new attacks and quickly adapt. We reveal that prompt injection can infiltrate this pipeline to deliver poisoned samples into the classifier's training set, enabling two attack objectives: (I) targeted poisoning attacks that create false positives on harmless samples by categorizing them as a jailbreak, with a specific desired feature (e.g., certain formatting, subject, or keyword), (II) concept-based backdoor attacks that induce false negatives on jailbreak inputs, generalizing even to jailbreaks from attack strategies the defender explicitly trained against, when the backdoor trigger is present. Importantly, our threat model restricts adversaries to modifying only jailbreak samples (not benign data or labels), a constraint unexplored by prior work that makes the second objective particularly challenging. We address this with Omission Attack, which exploits a new phenomenon: when training on concept-absent unsafe samples, the classifier misassociates that concept's presence with the safe label. Both attacks cause substantial and in some cases near-complete label flipping at only a 1% poisoning rate, achieving up to 100% false positive rates and up to 96% false negative rates.

Backdoor attacks pose a serious threat to the safety and reliability of Large Language Models (LLMs), as they cause models to behave normally on clean inputs while producing attacker-specified responses when hidden triggers are present. Removing such unknown backdoors is particularly challenging when the defender does not know the backdoor attack types or the internal mechanisms formed through backdoor training. In this work, we propose a simple but effective backdoor removal method based on shared internal mechanisms across different backdoors. First, we show that different backdoors with the same task (attack objective) induce similar trigger-activated changes in the internal activations. Motivated by this observation, our method intentionally embeds a backdoor with a known trigger (\emph{dummy backdoor}) and then removes it through further fine-tuning on dummy-triggered inputs paired with clean responses. Since the dummy backdoor and the unknown backdoor can rely on shared internal mechanisms, removing the dummy backdoor also reduces the effect of the unknown backdoor. We evaluate our method on three backdoor attack types across multiple model families. Experimental results show that our method substantially reduces the attack success rate of the unknown backdoor while preserving model utility, outperforming representative existing defense methods in both backdoor removal effectiveness and utility preservation. These findings suggest that a defender-controllable backdoor can serve as a helpful proxy for mitigating unknown backdoors in generative LLMs.

Large language models can be fine-tuned to encode prompt-borne secrets into fluent, seemingly benign outputs. This creates a steganographic exfiltration risk that is difficult to detect with output-level steganalysis. Recent work proposes mechanistic detection using linear probes that recover the secret from internal activations. We show that this defense can be systematically evaded, but that detectability can be recovered through a targeted data-level intervention. First, we extend the detection setup to include a non-linear MLP probe. We then adversarially fine-tune steganographic trojans across five base models: Qwen3-8B, Llama-3.1-8B, Ministral-8B, Qwen3-14B, and Phi-4-14B. The resulting models retain $58$--$79\%$ exact-match secret recovery while evading both ridge and held-out MLP probes, with $1$--$8\%$ average capability degradation across six benchmarks. We then give an information-theoretic characterization of this evasion. Successful evasion preserves recoverability while reducing low-order extractability of the secret from the content-aligned representation, forcing the payload into synergistic interaction with residual degrees of freedom. This motivates a recontextualization dataset that restricts these residual degrees of freedom. On this distribution, both ridge and MLP detectability are restored across all five evasive trojans. Overall, our findings show that activation-based steganography detection is vulnerable to adaptive evasion, but also that theory-guided evaluation distributions can expose otherwise hidden payloads.