Latent-space monitors aim to detect undesirable behaviours in Large Language Models by leveraging their internal representations rather than relying solely on black-box outputs. These methods have shown promise in identifying behaviours such as deception and unsafe completions. However, these monitors may themselves become training signals, for example, by using problematic samples found in deployment to retrain models. This raises an important question: can models learn to evade such monitors? To evaluate this capability, we introduce RL-Obfuscation, in which LLMs are finetuned via reinforcement learning to evade latent-space monitors while maintaining their blackbox behaviour. We apply RL-Obfuscation to Language Models ranging from 7B to 14B parameters and evaluate their Evasion Success Rate against a suite of monitors. We find that token-level monitors are highly vulnerable to this attack while more holistic monitors, such as max-pooling or attention-based probes, remain robust. Moreover, for these vulnerable monitors, models trained to evade a single static monitor can generalise to evade other unseen monitors. We also find that the models can be trained to conditionally bypass
Monitoring AIs at runtime can help us detect and stop harmful actions. In this paper, we study how to efficiently combine multiple runtime monitors into a single monitoring protocol. The protocol's objective is to maximize the probability of applying a safety intervention on misaligned outputs (i.e., maximize recall). Since running monitors and applying safety interventions are costly, the protocol also needs to adhere to an average-case budget constraint. Taking the monitors' performance and cost as given, we develop an algorithm to find the best protocol. The algorithm exhaustively searches over when and which monitors to call, and allocates safety interventions based on the Neyman-Pearson lemma. By focusing on likelihood ratios and strategically trading off spending on monitors against spending on interventions, we more than double our recall rate compared to a naive baseline in a code review setting. We also show that combining two monitors can Pareto dominate using either monitor alone. Our framework provides a principled methodology for combining existing monitors to detect undesirable behavior in cost-sensitive settings.
As autonomous agents become more capable of performing real-world tasks, distinguishing scheming behavior from benign task pursuit may become a central AI control problem. Existing monitors often rely on chain-of-thought access or internal activations, or use prompted frontier models, all of which can be unavailable, unreliable or expensive in deployment. In this work, we study action-only deliberative monitors: smaller open-weight models trained to detect scheming and sabotage from agentic trajectories without accessing the monitored agent's reasoning or model internals. Our method, inspired by deliberative alignment, uses a scheming specification to elicit structured rationales from a frontier teacher, filters them with a separate judge, and distills the highest-quality rationales into open-weight monitors with supervised fine-tuning and reinforcement learning. We train on five datasets, and evaluate across six out-of-distribution agentic misalignment benchmarks. We show that applying our method to Qwen3.5-27B yields higher performance than all low-cost frontier models as prompted monitors (Gemini 3.1 Flash-Lite, GPT-5.4 Nano, and Claude Haiku 4.5) and than Gemini 2.5 Pro, while
Since autonomous coding agents generate complex behaviors at high-volume, we may want to use other LLMs to monitor actions to reduce the risk from dangerous misaligned behavior. To better understand the limitations of such monitors against the diverse attack strategies that a coding agent could use, we present SLEIGHT-Bench (Subtle Low-itEration Insight-Guided Harmful Transcripts), a benchmark of synthetic transcripts containing 40 attacks across 11 categories, each showing a coding agent covertly pursuing a harmful objective (e.g. weight exfiltration, credential theft, unmonitored rogue deployment). We find that 20 of 40 attacks are never caught by an Opus 4.6 monitor with extended thinking across 10 trials at a 1% false-positive rate, while the overall catch rate is 32%. We identify several evasion strategies such as exploiting model priors, user instruction ambiguity, and state manipulation, which facilitate creating attacks that evade frontier monitors. We also elicit stronger monitor performance using coding agents as monitors versus regular prompted monitors, and for some evasion strategies show improved catch rates with targeted monitor prompts. Our dataset and evaluation fr
Activation monitors -- lightweight probes trained on a language model's internal representations -- are an increasingly common layer in deployment safety stacks. Deployed models however are rarely static: they are quantized, fine-tuned, adapted with LoRA, or served with merged adapters while the monitor remains frozen. We present the first systematic test of whether this implicit contract holds: whether activation monitors trained on a base model remain reliable after these routine model updates. Across multiple safety-relevant monitors, model depths, update families, and open-weight models, we find a sharp split: quantization-style updates largely preserve frozen probe performance, while fine-tuning-style updates frequently make probes stale. Fragility is highly monitor-dependent, with privacy/PII probes most affected and refusal-compliance probes comparatively stable, showing that retraining a behavior need not stale its corresponding monitor. QLoRA is especially damaging despite NF4 quantization alone being relatively benign, suggesting that quantization becomes riskier when combined with adaptation. We further show that degradation is predictable from pre-deployment features, e
Behavior Trees (BTs) are high level controllers that have found use in a wide range of robotics tasks. As they grow in popularity and usage, it is crucial to ensure that the appropriate tools and methods are available for ensuring they work as intended. To that end, we created a new methodology by which to create Runtime Monitors for BTs. These monitors can be used by the BT to correct when undesirable behavior is detected and are capable of handling LTL specifications. We demonstrate that in terms of runtime, the generated monitors are on par with monitors generated by existing tools and highlight certain features that make our method more desirable in various situations. We note that our method allows for our monitors to be swapped out with alternate monitors with fairly minimal user effort. Finally, our method ties in with our existing tool, BehaVerify, allowing for the verification of BTs with monitors.
Runtime monitors assess whether a system is in an unsafe state based on a stream of observations. We study the problem where the system is subject to probabilistic uncertainty and described by a hidden Markov model. A stream of observations is then unsafe if the probability of being in an unsafe state is above a threshold. A correct monitor recognizes the set of unsafe observations. The key contribution of this paper is the first correct-by-construction synthesis method for such monitors, represented as finite automata. The contribution combines four ingredients: First, we establish the coNP-hardness of checking whether an automaton is a correct monitor, i.e., a monitor without misclassifications. Second, we provide a reduction that reformulates the search for misclassifications into a standard probabilistic system synthesis problem. Third, we integrate the verification routine into an active automata learning routine to synthesize correct monitors. Fourth, we provide a prototypical implementation that shows the feasibility and limitations of the approach on a series of benchmarks.
Discrete diffusion language model (DLM) fine-tuning inherits inexpensive diagnostics from denoising-time confidence monitors, but their PEFT-training meaning is untested. We test top-1 argmax concentration as a collapse warning. Across 816 LoRA/PEFT configurations from three DLM families, the warning fires for every configuration while logs record 0/816 actual collapses at the 200 step horizon, giving zero precision. The cause is pre-equilibrium saturation: top-1 concentration is already high before optimization and quickly becomes insensitive to final training stability. We then evaluate max LoRA gradient norm, a parameter-side signal that samples gradient routing rather than token concentration. On a pooled held-out LLaDA-family split, a train-optimized threshold identifies top-decile final-loss configurations with precision 0.68 and F1=0.79, above the all-positive top-1 baseline even at the lower split-bootstrap confidence bound. Autoregressive controls and cross-family threshold failures bound the result to short-horizon DLM-LoRA inspection rather than a universal collapse detector. Workflow: drop top-1 as a PEFT alarm, log max-gradient early in training, and calibrate threshol
This paper focuses on the runtime verification of hyperproperties expressed in Hyper-recHML, an expressive yet simple logic for describing properties of sets of traces. To this end, we consider a simple language of monitors that observe sets of system executions and report verdicts w.r.t. a given Hyper-recHML formula. We first employ a unique omniscient monitor that centrally observes all system traces. Since centralised monitors are not ideal for distributed settings, we also provide a language for decentralized monitors, where each trace has a dedicated monitor; these monitors yield a unique verdict by communicating their observations to one another. For both the centralized and the decentralized settings, we provide a synthesis procedure that, given a formula, yields a monitor that is correct (i.e., sound and violation complete). A key step in proving the correctness of the synthesis for decentralized monitors is a result showing that, for each formula, the synthesized centralized monitor and its corresponding decentralized one are weakly bisimilar for a suitable notion of weak bisimulation.
As deploying large amounts of monitoring equipment results in elevated cost and power consumption, novel low-cost monitoring methods are being continuously investigated. A new technique called Power Profile Monitoring (PPM) has recently gained traction thanks to its ability to monitor an entire lightpath using a single post-processing unit at the lightpath receiver. PPM does not require to deploy an individual monitor for each span, as in the traditional monitoring technique using Optical Time-Domain Reflectometer (OTDR). In this work, we aim to quantify the cost and power consumption of PPM (using OTDR as a baseline reference), as this analysis can provide guidelines for the implementation and deployment of PPM. First, we discuss how PPM and OTDR monitors are deployed, and we formally state a new Optimized Monitoring Placement (OMP) problem for PPM. Solving the OMP problem allows to identify the minimum number of PPM monitors that guarantees that all links in the networks are monitored by at least $n$ PPM monitors (note that using $n>1$ allows for increased monitoring accuracy). We prove the NP-hardness of the OMP problem and formulate it using an Integer Linear Programming (IL
Current LLM safety defenses fail under decomposition attacks, where a malicious goal is decomposed into benign subtasks that circumvent refusals. The challenge lies in the existing shallow safety alignment techniques: they only detect harm in the immediate prompt and do not reason about long-range intent, leaving them blind to malicious intent that emerges over a sequence of seemingly benign instructions. We therefore propose adding an external monitor that observes the conversation at a higher granularity. To facilitate our study of monitoring decomposition attacks, we curate the largest and most diverse dataset to date, including question-answering, text-to-image, and agentic tasks. We verify our datasets by testing them on frontier LLMs and show an 87% attack success rate on average on GPT-4o. This confirms that decomposition attack is broadly effective. Additionally, we find that random tasks can be injected into the decomposed subtasks to further obfuscate malicious intents. To defend in real time, we propose a lightweight sequential monitoring framework that cumulatively evaluates each subtask. We show that a carefully prompt engineered lightweight monitor achieves a 93% defe
The behavior of neural networks (NNs) on previously unseen types of data (out-of-distribution or OOD) is typically unpredictable. This can be dangerous if the network's output is used for decision-making in a safety-critical system. Hence, detecting that an input is OOD is crucial for the safe application of the NN. Verification approaches do not scale to practical NNs, making runtime monitoring more appealing for practical use. While various monitors have been suggested recently, their optimization for a given problem, as well as comparison with each other and reproduction of results, remain challenging. We present a tool for users and developers of NN monitors. It allows for (i) application of various types of monitors from the literature to a given input NN, (ii) optimization of the monitor's hyperparameters, and (iii) experimental evaluation and comparison to other approaches. Besides, it facilitates the development of new monitoring approaches. We demonstrate the tool's usability on several use cases of different types of users as well as on a case study comparing different approaches from recent literature.
A monitor is a widely-used concurrent programming abstraction that encapsulates all shared state between threads. Monitors can be classified as being either implicit or explicit depending on the primitives they provide. Implicit monitors are much easier to program but typically not as efficient. To address this gap, there has been recent research on automatically synthesizing explicit-signal monitors from an implicit specification, but prior work does not exploit all paralellization opportunities due to the use of a single lock for the entire monitor. This paper presents a new technique for synthesizing fine-grained explicit-synchronization protocols from implicit monitors. Our method is based on two key innovations: First, we present a new static analysis for inferring safe interleavings that allow violating mutual exclusion of monitor operations without changing its semantics. Second, we use the results of this static analysis to generate a MaxSAT instance whose models correspond to correct-by-construction synchronization protocols. We have implemented our approach in a tool called Cortado and evaluate it on monitors that contain parallelization opportunities. Our evaluation show
Monitors are a key tool in the field of runtime verification, where they are used to verify system properties by analyzing execution traces generated by processes. Work on runtime monitoring carried out in a series of papers by Aceto et al.$~$has specified monitors using a variation on the regular fragment of Milner's CCS and studied two trace-based notions of equivalence over monitors, namely verdict and $ω$-verdict equivalence. This article is devoted to the study of the equational logic of monitors modulo those two notions of equivalence. It presents complete equational axiomatizations of verdict and $ω$-verdict equivalence for closed and open terms over recursion-free monitors. It is also shown that verdict equivalence has no finite equational axiomatization over open monitors when the set of actions is finite and contains at least two actions.
As AI systems are increasingly deployed in autonomous agentic settings at scale, it is important to ensure the actions they take are safe and aligned with user intent. Monitoring agent actions is a key safety mechanism, yet reliable monitors remain difficult to build and the scale of these systems makes human oversight impractical. We show that combining signals from diverse monitors into an ensemble improves detection of misaligned actions. We build 12 GPT-4.1-Mini monitors using both prompting and fine-tuning strategies. We evaluate them on coding tasks where candidate solutions pass standard tests but fail on adversarial inputs. In this setting, diverse ensembles outperform both individual monitors and homogeneous ensembles. Our best 3-monitor ensemble achieves 2.4x greater detection performance gain compared to an ensemble composed of three identical monitors, with the same ensemble performing strongly on an independent dataset. We contend that these results show that diversity - not scale - drives gains. The best ensembles combine strong individual performance with low correlation between monitors. Furthermore, fine-tuned monitors appear in every top-performing ensemble and ma
We present a model-based approach to learning robust runtime monitors for autonomous systems. Runtime monitors play a crucial role in raising the level of assurance by observing system behavior and predicting potential safety violations. In our approach, we propose to capture a system's (stochastic) behavior using interval Hidden Markov Models (iHMMs). The monitor then uses this learned iHMM to derive risk estimates for potential safety violations. The paper makes three key contributions: (1) it provides a formalization of the problem of learning robust runtime monitors, (2) introduces a novel framework that uses conformance-testing-based refinement for learning robust iHMMs with convergence guarantees, and (3) presents an efficient monitoring algorithm for computing risk estimates over iHMMs. Our empirical results demonstrate the efficacy of monitors learned using our approach, particularly when compared to model-free monitoring approaches that rely solely on collected data without access to a system model.
We introduce a red-teaming methodology that exposes harder-to-catch attacks for coding-agent monitors, suggesting that current practices may under-elicit attacks and overstate monitor performance. We identify three challenges with current red-teaming. First, mode collapse in attack generation, which we reduce with a novel attack taxonomy for broader coverage. Second, a conceive-execute gap: frontier LLMs can propose strong attack ideas or execute them, but not all at once. We mitigate this by decomposing attack construction into strategy generation, execution, and post-hoc trajectory refinement. Third, manual elicitation is costly to scale, which we address with our semi-automated red-teaming pipeline. Applied to BashArena, an AI control setting for tool-using coding agents, this pipeline produces MonitoringBench, a benchmark of 2,644 attack trajectories for evaluating monitor capabilities and failure modes. Our pipeline produces more diverse and stronger attacks: Opus-4.5 monitor's catch rate falls from 94.9\% on elicited-only Opus attacks to 60.3\% on our best refined attacks, with larger drops for several mid-tier monitors. Attacks optimized against three development monitors ge
Specifying informative and dense reward functions remains a pivotal challenge in Reinforcement Learning, as it directly affects the efficiency of agent training. In this work, we harness the expressive power of quantitative Linear Temporal Logic on finite traces (($\text{LTL}_f[\mathcal{F}]$)) to synthesize reward monitors that generate a dense stream of rewards for runtime-observable state trajectories. By providing nuanced feedback during training, these monitors guide agents toward optimal behaviour and help mitigate the well-known issue of sparse rewards under long-horizon decision making, which arises under the Boolean semantics dominating the current literature. Our framework is algorithm-agnostic and only relies on a state labelling function, and naturally accommodates specifying non-Markovian properties. Empirical results show that our quantitative monitors consistently subsume and, depending on the environment, outperform Boolean monitors in maximizing a quantitative measure of task completion and in reducing convergence time.
Monitoring the chain-of-thought (CoT) of reasoning models is a promising approach for detecting covert misbehavior (i.e., hidden objectives) in code generation tasks. While large models (GPT-5, Gemini-3-Flash) can serve as effective CoT monitors, they are expensive to deploy due to the lengthy reasoning traces and high API cost, emphasizing the need for smaller, cheaper alternatives. Nevertheless, we find that current small models (4B--8B) struggle to detect hidden objectives despite access to the CoT, frequently misattributing them as part of the user query. To address this, we propose a post-training pipeline combining supervised fine-tuning (SFT) and reinforcement learning (RL), where SFT narrows the gap for in-domain tasks by distilling detection behavior from stronger monitors, and RL on hard and subtly crafted hidden objectives helps the model generalize to out-of-domain monitoring tasks. To validate this generalization, we evaluate under a realistic threat model motivated by practical supply-chain attacks, where the adversary is a third-party LLM router injecting hidden objectives into code-generation requests through either prompt manipulation or code manipulation attacks.
Stream-based monitoring is a real-time safety assurance mechanism for complex cyber-physical systems such as unmanned aerial vehicles. In this context, a monitor aggregates streams of input data from sensors and other sources to give real-time statistics and assessments of the system's health. Since monitors are safety-critical components, it is crucial to ensure that they are free of potential runtime errors. One of the central challenges in designing reliable stream-based monitors is to deal with the asynchronous nature of data streams: in concrete applications, the different sensors being monitored produce values at different speeds, and it is the monitor's responsibility to correctly react to the asynchronous arrival of different streams of values. To ease this process, modern frameworks for stream-based monitoring such as RTLola feature an expressive specification language that allows to finely specify data synchronization policies. While this feature dramatically simplifies the design of monitors, it can also lead to subtle runtime errors. To mitigate this issue, this paper presents pacing types, a novel type system implemented in RTLola to ensure that monitors for asynchrono