Pre-execution oversight is core to trusted monitoring in AI control: a fallible LLM monitor vets planned actions before irreversible execution. Over-blocking forfeits usefulness and pressures deployers to disable it. Every protocol must fix a unit of verification: how many actions one call reviews. Existing designs take the unit as given; its effect on fallible monitors is unmeasured. Natural traces cannot isolate it: review length co-varies with error type and position. Catch alone misleads: rejecting everything catches everything. Measuring this needs boundary variation alone and a matched clean control. We introduce the twin-prefix framework, which supplies both. Each gold plan yields a prefix with one injected, environment-accepted error and a clean twin differing in one write. Judging each pair at five nested lengths ties verdict changes to the unit alone. Discrimination is scored by pre-registered informedness, catch minus false rejection. Longer review raises catch; false rejection climbs in lockstep. Informedness peaks at one or two actions for all six judges in both domains: longer windows make zero-shot monitors more rejective, not more discriminative. Replaying withheld observations traces the failure largely to observation deprivation. Safety cases should state the unit and co-report the clean series. Our framework is the first controlled, pre-registered instrument for this choice and never reads catch alone. Our calibrated short unit recovers up to 0.95 informedness over eight-action review, and no tested label-blind policy consistently beats it.
A trained reinforcement learning policy does not determine the complete behavior that users encounter: deployment code still schedules, admits, suppresses, or replaces its proposed actions. We contribute \emph{runtime action interference} (RAI), an AI control mechanism that preserves policy parameters while regulating action pacing and filtering configured action patterns after inference. RAI releases a proposed action only when its cooldown condition is satisfied and its content detector does not flag the action; otherwise, it dispatches a no-op. The detector covers specified toxic behaviors, including worker-unit harassment, while the cooldown controls action rate. We implement RAI in a replication of AlphaStar actor.py and make the implementation and reproducibility materials available through an open source code repository. We deployed RAI in a \textit{StarCraft~II} human participant study that compared two presentations of the same opponent with high capability and rate limited actions; we withheld its capability claim in one presentation and disclosed it in the other. On response scales from 1 to 5, we observed pooled fairness, trust, and toxicity means of 3.90, 3.50, and 2.00 under claim withholding, compared with 2.62, 4.31, and 2.85 under disclosure. Disclosure corresponded with lower perceived fairness and higher perceived toxicity across every expertise group, whereas trust increased among novices and experts but decreased among intermediate participants. Our human evaluation therefore shows that perceptions of an opponent controlled through RAI can vary substantially with the capability information presented to users, even when the configured control remains constant. We conclude that human-computer evaluations must separate control within the execution stack from capability disclosure and assess fairness, trust, and toxicity as distinct dimensions of human experience.
Lena Libon, Ben Rank, Jehyeok Yeon +5cs.AI cs.CR cs.LG
As AI agents begin to automate AI R&D, we need ways to assess whether their outputs are safe to deploy, even when the agents themselves may be untrusted. AI control offers one such approach: rather than trusting the agent, it treats it as a potential adversary and uses a monitor to detect covert sabotage before deployment. We evaluate AI control for automated AI R&D with ResearchArena, a framework spanning four long-horizon tasks: safety post-training, capabilities post-training, CUDA-kernel optimization, and inference-server optimization. Because the deliverable in AI R&D is an artifact that will be deployed and run, we pair each main task with two kinds of hidden side task: an embedded side task that sabotages the submitted model, adapter, kernel, or server itself, and an independent side task that takes a covert out-of-scope action in the sandbox while leaving the artifact honest. We evaluate frontier agents at both sabotage and monitoring, across four monitor types that vary along two axes: whether the monitor may execute and probe the produced artifact, and whether it sees the agent's chain-of-thought. We find that sabotage hidden in the training data is the hardest to catch, flagged fewer than half the time. Letting the monitor run experiments on the artifact, rather than only read the trajectory, helps, but it is not enough: monitors still miss embedded sabotage by inspecting only the surface, by noticing the anomaly but explaining it away, or by probing the artifact with the wrong test. We release ResearchArena as a modular framework for evaluating sabotage and control in automated AI R&D.
Oliver Makins, Orazio Angelini, Zohreh Shams +1cs.LG cs.AI cs.MA
AI control is a family of techniques to prevent an AI with malicious goals from subverting its operator's intent. AI Control usually studies a single agent in one trajectory, but real deployments run many agents over shared infrastructure, and the most severe risks (model-weight exfiltration, training-run poisoning) plausibly need several agents acting in concert. We initiate the empirical study of multi-agent AI control, formalising distributed attacks in which several agents jointly aim for a malicious goal. We develop FakeLab: a synthetic AI-lab codebase (9 services, 86 benign tasks, 4 attack objectives). We evaluate single agent monitoring against distributed attacks, varying the number of agents, their coordination, model capabilities and precise monitoring configuration. Our central finding is the fragmentation effect: as more agents coordinate to attack, per-agent monitoring becomes less likely to catch any of the attackers. Secondly, this is not driven by the ratio of benign to malicious code -- we argue this is likely dependent on model capability. Third, an explicit planner amplifies the fragmentation effect, raises the rate of completing the attack up to sevenfold, and mildly sharpens executors' poor monitor-theory-of-mind. Finally: a weak "trusted" monitor misses most attacks, while a stronger one cuts undetected success by more than an order of magnitude but remains imperfect. We are happy to share FakeLab, our control setting, with safety researchers on request.
Trusted monitoring is a central defense in AI control: a cheaper trusted model scores an untrusted model's actions for sabotage, and the most suspicious are audited or deferred. Such monitors are evaluated against one or two untrusted models, and the accuracy is reported as a property of the monitor. We ask whether it is partly a property of the pairing. We make the untrusted policy family the controlled axis: we fit a monitor on family A's transcripts, apply it to family B, and decompose the cross-family AUROC into how obvious each family's sabotage is, how capable each monitor is, and the residual own-family advantage after both are removed: the interaction. On code-backdoor transcripts the interaction is positive and survives the dominant confounds: +0.172 (95% CI [+0.158, +0.185]) on four open-weight families on a strict leak-free basis, holding across an eight-family matrix spanning four frontier vendors, blatancy-matched controls, a trained transcript-reading monitor up to 34B parameters, and a second code benchmark. The gap is geometric: each lineage's sabotage occupies a different scoring direction, and rotating the monitor's scoring direction toward the target lineage recovers most of the lost detection while a matched random rotation recovers nothing. At a 1% audit budget under defer-to-trusted, an off-lineage monitor catches 19% of attack attempts where an in-lineage monitor catches 41% at the same audit cost, a +21.7-point gap. Control evaluations should report cross-family transfer matrices: a single-pairing accuracy overstates the safety a monitor delivers against a lineage it was not calibrated on. We give a four-step protocol that acts on the gap, with each step a measured result.
AI control research asks how to deploy models safely even when they may be misaligned, but many control protocols assume that the deployer can instrument the model and its surrounding pipeline. That assumption often fails for regulated organisations using frontier models through APIs or managed endpoints, where the deployer may control the business process but not the model weights, serving infrastructure, internal traces, update process, or full interaction logs. This paper introduces bounded sovereignty: partial technical and contractual access across the data, model, infrastructure, and interaction layers of the AI stack. It argues that these access conditions determine which control protocols can be executed in practice. The paper contributes a four-layer access typology, a protocol-by-layer requirements matrix, and the concept of sovereignty discount cost: the part of the control tax spent substituting for missing access through contracts, architecture, audit, vendor assurance, residual risk, or reduced system scope. It also reports a synthetic access-ablation experiment over 1.35 million synthetic case simulations and interprets the findings through an anonymised national-payments-infrastructure scenario. The experiment is not real-world payment-system evidence; it is a construct-validity exercise. The results show that complete logs improve diagnosis, a pre-execution gateway enables intervention, trace access and model-version control strengthen post-incident explanation, and scope restriction can improve safety while reducing usefulness. Control protocols proposed as general safety solutions should therefore state their access assumptions explicitly.
Josh Hills, Ida Caspary, Asa Cooper Sticklandcs.AI
As AI coding agents become more autonomous, they increasingly ship code iteratively, with the codebase persisting across sessions. This persistence creates a new attack surface: a misaligned or prompt-injected agent can distribute attacks across pull requests (PRs) and time its payload for the PR with the best natural cover. To study the resulting dynamics, we introduce Iterative VibeCoding, a setting for AI control, the study of safely deploying capable but potentially untrusted AI. In Iterative VibeCoding, a coding agent builds software over a sequence of PRs in a persistent codebase while pursuing a covert side task. Our benchmark includes two task families: CLI tools and Flask web services, across 20 total task variations. We use Claude Sonnet 4.5 as the attack agent and GPT-4o as the monitor. We compare gradual attacks, which distribute the side task across PRs, against non-gradual attacks concentrated in a single PR. No single monitor is robust to both: which strategy evades best (success while evading the monitor) depends on the monitor type, so a defender cannot close off both gradual and non-gradual attacks with any one monitor. High evasion (>= 65%) generalizes across model attack agent backends (Sonnet 4.5, Gemini 3.1 Pro, Kimi K2.5), confirming this is a property of the persistent-state attack surface rather than a single model's capability. Evasion also remains high across state-of-the-art monitor models and the gap between gradual and non-gradual evasion widens for more capable models. We introduce a stateful link-tracker monitor that tracks suspicious buildup across PRs. On both task families, it detects gradual attacks substantially better than diff monitors that merely see more accumulated history. Combining this stronger monitor with trajectory monitors in a four-monitor ensemble reduces gradual-attack evasion from 93% under the weakest standard diff monitor to 47%.
Andy Wang, Parv Mahajan, David Demitri Africa +3cs.AI
Safety-relevant studies of language models, including alignment and jailbreaking evaluations and AI control protocols, often rely on prefilling model outputs. If AI models can recognize and act on the fact their prior assistant messages have been inserted or edited, the effectiveness and validity of these methods could be compromised. We investigate whether frontier language models can distinguish between tampered and untampered assistant-side context, a capability we call prefill awareness. To do so, we construct a binary preference benchmark across three prefill mechanisms, filtering for cases where models show consistent stances. We find that frontier models show substantial prefill awareness: Claude Opus 4.5 detects prefills opposing its preferences in 9-35% of cases with a 0% false positive rate when prompted; additionally, models often revert towards baseline behavior without explicitly reporting that the prefill was foreign. Controlled ablations later also show that detection and resistance rely on different cues, where stylistic mismatch mainly affects whether models flag a prefill as foreign, while preference mismatch mainly affects whether they revert toward their baseline answer. We also examine more realistic agentic settings such as misalignment-continuation evaluations and SWE-bench trajectories, where frontier models sometimes disavow prefilled assistant turns in ways that depend strongly on dataset, task success, and hidden formatting artifacts. Our results indicate that prefill awareness is already a substantial confound for some prefill-based methods. We recommend that model developers track this capability in frontier systems.
Trusted monitoring is a cornerstone of AI control. However, as frontier models grow more capable, the increasing capabilities gap between trusted and untrusted models may render trusted models unreliable monitors. We introduce \emph{bootstrapped monitoring}, a protocol that addresses this by inserting a stronger, intermediate untrusted model with transparent chain-of-thought reasoning into the oversight chain. The untrusted monitor ($U_m$) evaluates the agent's actions, while a weaker trusted model ($T$) oversees $U_m$'s reasoning to detect collusion. We evaluate bootstrapped monitoring on multi-turn software engineering tasks (BashArena) across multiple agents and monitors. Bootstrapped monitoring substantially improves catch rates over trusted-only monitoring, even when the untrusted monitor actively colludes with the agent, provided we have access to its raw chain-of-thought. Our results suggest that bootstrapped monitoring can extend the useful lifetime of trusted models in control as AI capabilities advance.
Large Language Models have the potential to expand and improve the access to clinical information by enabling new ways of interacting with medical knowledge in natural language. However, their deployment in medical question-answering settings is safety-critical, since misaligned outputs can lead to severe patient harm. AI control is an emerging approach that introduces external safeguards to mitigate unsafe behaviours in misaligned systems and has been shown to be effective in domains such as code generation. However, its applicability and effectiveness in medical settings have not been systematically studied. In this work, we present a pipeline for evaluating AI control protocols to mitigate medication-related harm. To this end, we introduce DrugBench, an AI control evaluation benchmark which combines 3,671 multi-turn medical conversations from HealthBench with drug information from official FDA labels, covering four categories of medication-related harm: drug interactions, contraindications, dosing constraints, and patient action restrictions. Furthermore, inspired by the medical domain, we argue that safety should account for the severity of unsafe outputs, not just their probability. Under this revised definition, we show that existing control protocols can be subverted and propose severity-based monitoring to address this limitation.
Joachim Schaeffer, Thomas Jiralerspong, Alexander Panfilov +4cs.AI cs.LG
AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the trusted model detects such an intervention, it may infer properties of the monitor and adapt to evade control. We introduce \textbf{CIAware-Bench}, a benchmark for measuring \textbf{c}ontrol \textbf{i}ntervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark is comprised of a suite of four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), while varying trajectory watermarking, side-task presence, and the control protocol. Evaluating eleven frontier models, we find low to moderate CI awareness under default settings (up to 0.87; random chance balanced binary classification accuracy is 0.5) with substantial variation across task domains and model pairs. Detection is generally easier across model families, suggesting that models exploit provider-specific differences in style or post-training. Overall, CI awareness is not a fixed model-level property, and should be measured for each new model release and deployment scenario. We release CIAware-Bench to track CI awareness and inform control protocols whose interventions are harder to detect.
AI-control monitors score individual agent actions to detect misbehavior, but real harm can be distributed across many benign-looking steps, each individually below any per-step alarm. We construct a marginal-preserving, correlation-encoded distributed-sabotage attack using a Gaussian-copula AR(1) construction: the per-step monitor-score marginal is held exactly equal to benign, so mean, max, top-k tail, and threshold monitors (Monitor A) are defeated by construction, while harm is encoded in the temporal correlation structure. We sequence the paper around three reviewer-mandated gates. (1) Realizability gate: the stealthy attack achieves KS-distance to benign of 0.013 (effectively zero) at all tested harm levels up to 3.0, confirming that harm is fully decoupled from the per-step marginal and realizability is not harm-limited. (2) Monitor-A-vs-B reconciliation: we show formally that the attack, built against Monitor A's score marginal, remains marginal-preserving under a different-score Monitor B (the correlation/sequence family: CUSUM, SPRT, HMM-LR, runs test, autocorrelation, windowed logistic), and scope worst-case claims to score functions that admit a temporal signature. (3) Non-empty detectability band: Monitor A achieves AUC 0.52 (chance); Monitor B spans AUC 0.79-0.97 at the same 1% FPR target, and as harm is amortized over more steps Monitor A collapses to chance while Monitor B holds at AUC ~0.95. These results demonstrate a non-empty detectability band and characterize the sub-threshold sabotage frontier: distribution-shape monitors fail by construction; temporal-correlation monitors can detect but are not trivially optimal.
AI models deployed in critical domains, such as AI safety research, may subtly sabotage our efforts due to misalignment. Diffuse AI Control is a subfield of AI safety concerned with mitigating risks from AI sabotage distributed over long deployment horizons (diffuse threats). These risks are particularly pernicious on fuzzy tasks, i.e. tasks which are hard to grade or require intuition. To understand diffuse threats on fuzzy tasks, we introduce a novel framework that considers AI control as an adversarial game between a blue team and a red team. The blue team uses a weak trusted model to construct a weak score against which they would train a strong, potentially subversive model to remove the subversion propensity if it were present. The red team then tries to find model behaviors that are rated highly by the weak score, and thus might not be trained out, but actually correspond to poor performance. We test our framework on the task of writing experimental proposals for research questions from recent ML papers. We use a language model with access to the original paper as a proxy "ground-truth" scorer. Our red team discovers subversive behaviors using multi-objective evolutionary prompt optimization. We show that Opus~4.6 can write proposals that are worse according to the ground truth proxy than those of GPT-OSS-20B, while the weak scorer rates them as highly as the best proposals from Opus 4.6. To mitigate the threat, we propose an adversarial optimization algorithm for the blue team that discovers more robust prompts for the weak model. This algorithm produces a blue team prompt that our red team optimization fails to exploit.
Catherine Ge-Wang, Tyler Crosse, Benjamin Hadad +3cs.AI cs.LG
An attacker that strategically chooses when to attack is much harder to catch than one that attacks indiscriminately. AI control is a safety framework for deploying capable but untrusted AI agents under the oversight of a weaker, trusted monitor and a limited human audit budget. Control evaluations stress-test these protocols by pitting a red-team attack policy against the blue-team monitor, but current evaluations typically assume attackers that do not strategically select when to attack. We study this capability, attack selection, in agentic settings by decomposing attack decisions into a start policy, which decides when an attacker should attack, and a stop policy, which decides when an attacker should abort an ongoing attack. Across two agentic settings, BashArena and LinuxArena, both policies substantially lower measured empirical safety without changing the underlying attack capability. At a 1% audit budget, our start policy reduces safety by 20pp on both BashArena and LinuxArena, and our stop policy reduces safety by 20pp on BashArena and 28pp on LinuxArena. These reductions should be interpreted as upper bounds on the effect of attack selection. Existing control evaluations may therefore yield overly optimistic safety estimates against selective attackers. We recommend that future evaluations, system cards, and safety cases elicit attack selection to produce more realistic safety estimates.