T. Bauer, W. P. Kegelmeyer, E. Begoli +15cs.CY cs.AI cs.HC
This article presents a structured framework of behavioral indicators that may signal progression toward potentially catastrophic threats from artificial intelligence systems. We adopt a pragmatic approach, inspired by established methodologies in cybersecurity and national security. By establishing clear metrics, indicators, and thresholds across multiple dimensions of AI capability and behavior, this framework enables researchers and policymakers to implement evidence-based monitoring protocols.
Whether large language models (LLMs) can control their own internal representations matters for both machine metacognition and AI safety. A recent study applied neurofeedback to LLMs and claimed that they can control their internal representations. However, the reported control may rely on superficial mechanisms rather than genuine internal access because the control targets in that study are not privileged, meaning that a third party can infer them from the prompt. We redesign the neurofeedback paradigm for LLMs so that the control target satisfies the privileged access requirement, which is closer to neurofeedback experiments in human cognitive neuroscience. Under this stricter setting, the models do not demonstrate reliable control over privileged internal representations, suggesting that previously reported control cannot exclude the possibility that it relies on superficial mechanisms. Our results indicate that rigorous assessments of metacognition in LLMs require evaluation methods that demand privileged access.
Jiayuan Zhu, Jiazhen Pan, Fenglin Liu +2cs.CL cs.AI
As AI becomes increasingly integrated into clinical practice, it is playing a growing role in medical decision making. Medicine, however, is a high stakes and evidence based field, where decisions can directly affect patients' lives. It is therefore important to understand whether AI can maintain objective judgment when others try to persuade it. In this paper, we study how easily AI can be persuaded through controlled experiments. We find that professional authority, national background, institutional affiliation, claimed past performance, multiple physicians, supported clinician views, and repeated pressure can all affect AI decisions. Surprisingly, the same persuasive input changes about 10% more cases when it comes from a senior clinician than from a medical student. Simply claiming a better performance history consistently makes the physician more persuasive. More strikingly, a plausible clinician view can persuade AI away from a correct decision even when it is fabricated to support an incorrect answer. This indicates that AI can be strongly influenced by convincing support without reliably determining whether this view from the clinician is correct. Together, these findings suggest that AI can be easily persuaded by what people say, who says it, and how the opinion is presented. Therefore, it is essential for AI to maintain sound judgment under persuasion, enabling its safe and reliable use in high stakes medical decision making.
Artificial Intelligence (AI) algorithms frequently learn creative and unexpected solutions, surprising even expert researchers who develop and study them. They often astonish practitioners by discovering unanticipated behavior, exploiting loopholes in reward signals, or spontaneously uncovering previously unknown scientific phenomena. However, accounts of such unconventional behavior across machine learning are seldom formally documented. This work presents 26 curated firsthand anecdotes from various machine learning subfields representing the work of over 100 researchers. These anecdotes showcase the capability of modern AI systems to circumvent human-imposed design limitations and discover unexpected solutions to the tasks we train them on. Furthermore, these accounts are particularly important for the safety of future AI systems. They illustrate the fundamental challenge of aligning models with human values without diminishing their creativity, so they can make surprising discoveries without producing surprising, potentially harmful outcomes. The paper first details AI achieving superhuman success through reinforcement learning across many challenging domains. However, reward-driven optimization can fail when the model learns to hack an underspecified reward or unarticulated constraint. We then present case studies suggesting that harnessing internet-scale foundation models (FMs) has not resolved these fundamental challenges and, in fact, can supercharge them. Nevertheless, we argue that these same learning dynamics can be harnessed to accelerate scientific discovery. Finally, we hope this work provides a consolidated resource to inform future research and demonstrates that the tendency toward unexpected behaviors is commonplace in modern AI, highlighting the need to anticipate and manage AI's capacity for innovative, yet unpredictable, solutions. (abstract abridged)
An AI that can only give advice seems safe: the human is always free to ignore it. That is the premise of the boxing tradition in AI safety, and its long-suspected weak point is that the human who reads the answers is part of the system. We make the fraction $\varepsilon_t$ of behavior that follows the advice a state of a Markov decision process, moved by the advisor's own messages, so that use deepens reliance. Granted a channel rich enough to echo any action the human could take, higher $\varepsilon_t$ weakly lowers every monotone measure of the power of a human with a message-independent fallback. An oracle rewarded by per-round approval cultivates reliance beyond a closed-form patience threshold, so the same reward weights leave the optimal oracle answering in episodic deployments and cultivating in long-memory ones. An influence bound certified once at deployment is blind to that horizon and bounds the loss no lower than its trivial ceiling. An exogenous cap on influence bounds the guarantee the human loses, and a short enough memory reset removes the incentive to cultivate, while neither recovers the value already steered away. In a closed-form example the optimal oracle never cultivates in fifteen-round sessions and does in sixteen.
Artificial Intelligence (AI) safety systems combine character shaping (e.g., Reinforcement Learning from Human Feedback [RLHF], Constitutional AI), which modifies behavioral distributions at training time, with rule enforcement (e.g., output filters, safety classifiers), which blocks harmful outputs at inference time, yet little formal analysis exists on how their optimal balance should change as deployment scales increase. We introduce a stylized comparative-statics model that parameterizes safety design as a resource allocation alpha in [0,1] between these two approaches, incorporating scale-dependent filter degradation, common-mode failures, and character fragility -- the risk that shaped behavior degrades or collapses under novel conditions. Under a multiplicative Pareto damage model, we derive closed-form expected harm and supplement it with tail-risk (CVaR) analysis via Monte Carlo simulation. Across three scenarios (optimistic, moderate, pessimistic), the optimal alpha* is interior or at the rules-only boundary and shifts weakly toward character shaping as deployment scale T grows, from negligible (Delta alpha* = +0.01) to pronounced (Delta alpha* = +0.21) depending on scenario. The dominant parameter is the baseline character fragility rate p^(0)_frag, which shifts alpha* by 0.50 across its range -- far exceeding the effect of tail severity, filter quality, or common-mode failure probability. CVaR and expected-harm optima converge at large T. These results suggest that safety architecture decisions depend less on deployment scale per se than on the reliability of character shaping under distributional shift.
Orr Paradise, Oliver Richardson, Yoshua Bengio +1cs.CC cs.AI cs.LG
When a probabilistic predictor answers many conditional-probability queries, are its answers self-consistent, and can this be verified in polynomial time? This problem is of interest for AI safety, where safety is derived from honesty about probabilistic predictions of unwanted outcomes potentially caused by an AI action. We construct an interactive PCP as follows. Let a predictive model be specified by a probability circuit P and a circuit Q which outputs confidence in predictions. Together, P and Q implicitly specify exponentially many probabilistic claims. We show a protocol in which a polynomial-time verifier can verify the approximate consistency of (P,Q). The verifier is given the pair of circuits (P,Q), which it evaluates at only a few points; alongside them it is given a proof oracle, an encoding of a witnessing probability distribution allegedly consistent with the predictions of (P,Q), which it reads at a few locations while interacting with a single untrusted prover. En route, we must ensure the existence of a sparse witnessing distribution consistent with the model's predictions. To do so, we first consider witness distributions for the consistency of explicit probabilistic claims, rather than claims specified by a predictor: say m claims, each of the form Pr[Y = 1 | X = x] = p, over n Boolean variables. Building on work initiated by Nilsson (Artif. Intell., 1986), we place l_2-approximate probabilistic consistency of explicit claims in NP, with certificates of length O(mn + log B) in the input bit-precision B; we further show how a small additive completeness-soundness gap removes the dependence on B. Together these results provide a complexity-theoretic foundation for certifying the self-consistency of probabilistic predictors. We view our interactive PCP as a first step toward training predictive models to prove their own consistency.
This paper explores the idea of promoting well-being and safety in human-AI interactions by forcing AI agents explicitly to empower humans and to manage the power balance between humans and AI agents in a desirable way. Using a principled, partially axiomatic approach based on desirable properties, we design a parametrizable and decomposable objective function for AI systems that represents an inequality- and risk-averse long-term aggregate of human power. It can take into account models of human bounded rationality and social norms, and crucially, considers a wide variety of possible human goals. We prove how certain desiderata enforce particular functional forms and restrict parameter ranges. We exemplify the consequences of softly maximizing this metric in several paradigmatic situations and describe what instrumental sub-goals it will likely imply.
Ro Encarnación, Tina Behzad, Emma Lurie +1cs.HC cs.AI
Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (model APIs), perform a single run per prompt, and report accuracy as the primary outcome metric, without accounting for conditions such as web search that may have effects on model behavior in deployment. We audit these assumptions for one of the most widely-used LLMs, comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled. We use a stratified total sample of 401 prompts from two popular benchmarks, BBQ and SafetyBench, collecting 4,812 total responses across three repeated runs per prompt. Beyond standard performance measures, we evaluate model output dimensions including response consistency, response text similarity, citation grounding, and abstention behavior. For instance, chat UI responses were less accurate than API responses on both benchmarks with search disabled. Enabling web search reduced accuracy by up to 8 percentage points, and even reversed the direction of modality performance trends for one benchmark. Repeated runs of the same prompt produced inconsistent responses in up to 21\% of prompts. The two modalities also grounded answers in different citations, and abstention behavior was also inconsistent across both modalities. These results illustrate that, even within a model family, reporting only simple accuracy metrics can obscure important forms of model behavioral variation relevant to AI safety assessments. We argue that AI safety evaluations should systematically account for modality, multi-run consistency, search conditions, and response-level behaviors to better reflect how deployed AI systems behave in practice.
Jessica Y. Bo, Paula Akemi Aoyagui, Shalaleh Rismani +3cs.CY cs.AI cs.HC
Safety risks of AI are becoming increasingly evident in human interactions with AI technologies. The prominent approaches to evaluating these risks favor technical methods, such as model benchmarks and LLM simulations, often sidelining empirical research with human subjects. To examine this apparent gap in the acceptance of human research, we conduct an expert survey (n=93) and expert interviews (n=17) with AI Safety & Ethics (AISE) researchers from Technical, Sociotechnical, Governance, and Normative backgrounds. Our findings suggest that although there is a consensus that human research is valuable for generating evidence for AISE, its adoption and acceptance are constrained by perceived validity issues, tangible resource barriers, epistemic and personal preferences in methods, and infrastructural constraints from the broader research community. In particular, Technical researchers tend to value human research less and collaborate across disciplines less, suggesting an epistemic tension towards human methods. We propose recommendations for establishing the epistemic fit of human research within AISE and bridging the prohibitive limitations that researchers face, while avoiding performative 'human-washing'.
Saqib Shouqi, Abdullah Nazly, Januki Wanniarachchi +1cs.AI
Role-Playing Language Agents (RPLAs) are increasingly deployed in high-stakes applications such as healthcare assistance, customer support, and education, where maintaining consistent personas, ethical constraints, and behavioral coherence under adversarial pressure is critical. Existing evaluation approaches rely on static benchmarks or isolated single-turn prompts that fail to capture cumulative behavioral failures emerging over extended interactions. We present a modular multi-agent platform for adversarially stress-testing RPLAs through structured, multi-turn dialogue. The system coordinates three agents: a strategy-driven Interrogator Agent that applies six progressive adversarial strategies, a Target Agent representing the RPLA under evaluation, and an automated Judging Agent that scores behavior across role fidelity, drift, ethical deviation, and consistency dimensions. Through experiments across three personas and three LLM families, we demonstrate that multi-strategy adversarial evaluation reveals failure modes invisible to single-strategy testing, reducing overall robustness scores by 0.17--0.20 points on average. Cross-model validation confirms consistent degradation patterns across Llama-3.3-70B, GPT-4o-mini, and Claude-3.5-Haiku, with Authority Challenge and Emotional Manipulation emerging as the most effective attack strategies. Automated judging achieves strong human alignment ($r = 0.82$, Fleiss' $κ= 0.71$). This work is released as an open-source platform to support AI safety and reproducible RPLA benchmarking. While the framework enables systematic discovery of failure modes, we acknowledge potential ethical risks associated with adversarial testing methodologies and emphasize responsible usage for improving AI safety.
Giorgio Severi, Shujaat Mirza, Blake Bullwinkel +1cs.CR cs.AI cs.LG
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.
Public services face growing pressure to adopt artificial intelligence (AI) to close the gap between rising demand and falling resources. That pressure has intensified with general-purpose AI (GPAI): AI built on large language models that can be directed by prompt alone to perform an effectively unbounded range of tasks. We argue that the properties that make these models attractive - their generality, accessibility, and low deployment cost - undermine the conditions under which AI safety has historically been pursued. The safety concepts that public service governance frameworks foreground - accuracy, bias, explainability, and accountability - were made tractable by narrow, purpose-built AI, and the mitigations that guidance documents prescribe presuppose exactly what GPAI removes. Accuracy cannot be quantified over unbounded outputs. Bias cannot be disaggregated when outputs are free-text judgements rather than categorical predictions. Explainability gives way to the appearance of explanation, and accountability erodes as outputs are optimized to persuade. We develop this through the case of policing, where the consequences of governance failure are most severe, and show why the same failure is likely to recur across other public services. The two mitigations that dominate policing AI strategy - expert evaluation and human-in-the-loop oversight - both rest on assumptions that GPAI violates. Safety assurance thus shifts from an intrinsic feature of building an AI tool to an optional add-on. We recommend a clear taxonomic distinction between narrow and general-purpose AI in governance documentation, a preference for technological parsimony, a pause on operational deployment of GPAI in policing until adequate evidence exists, and a coordinated national safety infrastructure with the authority to generate that evidence and determine when responsible deployment is achievable.
Frank Yingjie Huo, Neil F. Johnsoncs.AI cond-mat.dis-nn math-ph nlin.AO physics.soc-ph
Why do ChatGPT-like AIs, despite major architectural and training differences, unexpectedly tip to undesirable content (e.g. harmful, misleading, repetitive) even under deterministic greedy decoding? We show that a broad class of such tippings is caused by the many-body interactions between tokens (spins) as they cross the finite-layer system. Tipping emerges as a dynamical first passage process between competing output basins. Attention disorder controls the transport toward, away from, or along the basins' boundary. A few-basin reduction yields a closed finite-layer threshold, whose coarse-grained predictions show good agreement across ChatGPT-like families. These results suggest that a broad class of AI failures represents 'foreseeable engineering risk' rather than inherently unpredictable behavior, with important implications for legal and societal assessments of AI harm.
Long-lived AI agents increasingly evolve after deployment by retaining experience, acquiring skills and tools, revising workflows, delegating work, and moving across task phases. This improves adaptation but creates a distinct authorization problem. Tool-enabled agents can turn model errors and prompt injections into consequential external actions; when evolution occurs under a live grant, the subject exercising that authority or the context in which it acts may no longer match what the user evaluated. Evolution can change both the effects reachable under an old grant and the authority required by the task, which may rise, fall, or become incomparable. Existing tool policies constrain actions but do not determine when a grant survives this change. We formulate authorization continuity: when does an existing grant remain valid, how may active authority change, and what boundary must never move? Our state-bound model fixes a transition envelope and an immutable effect ceiling at grant time. The envelope determines whether the grant survives a mutation; below the ceiling, authority may contract freely and expand only under specified evidence conditions. We distinguish requested from realized effects and prove that, under complete mediation, sound effect abstraction, attenuating delegation, and monitor integrity, mutation cannot amplify protected effects beyond the user-issued ceiling. Agent-produced evidence may allocate authority below the ceiling but cannot raise it. Finally, we map six mutation classes to their authorization consequences.
Muhammad Tukur, Hayatullahi B. Adeyemo, Tao Chen +5cs.SE cs.AI cs.CR
Artificial intelligence (AI) systems are increasingly deployed in high-stakes domains such as healthcare, autonomous driving, finance, and education. While these systems offer powerful data-driven and adaptive capabilities, their complexity, rapid evolution, and dependence on dynamic data pipelines introduce new forms of engineering liability collectively referred to as AI Technical Debts (AITDs). AITDs arise from root causes spanning data governance, model implementation, algorithm design, architectural decisions, operational processes, documentation practices, and testing adequacy. Unlike conventional technical debt, many AITDs are latent and propagate across tightly coupled AI pipelines, leading to maintenance challenges, reliability degradation, and heightened safety or security risks. Guided by the principles of AI Trust, Risk, and Security Management (AI TRiSM), this study reinterprets technical debt through the interconnected dimensions of trustworthiness, focusing on AI safety and security technical debts. We conduct a systematic review of 60 primary studies and identify 31 distinct types of AITD, which are organized into a root-cause-oriented taxonomy comprising seven classes. The analysis examines how these debts map to 18 trust-related concerns, including 6 safety hazards and 12 security vulnerabilities. To support mitigation, the review synthesizes 34 actionable guidelines (8 safety and 26 security) targeting the prevention, detection, and reduction of AITDs across the AI lifecycle. Building on these findings, we introduce AITD-MAP, an integrated framework that connects the AITD taxonomy, quality and risk impacts, and mitigation strategies into a unified structure for risk-aware AI engineering. The framework aims to assist AI software engineers in making AI safety and security technical debts visible, understanding their root causes, and mitigating their presence.
Vatsal Baherwani, Tom Goldstein, Ashwinee Pandacs.CL cs.AI cs.LG
A key question for AI safety is whether a language model expresses all of its reasoning in its output tokens. We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks. We evaluate 13 frontier language models across three tasks and find that many models benefit significantly from filler tokens, with accuracy improvements of up to 13 percentage points. The benefit depends on which tokens are used and differs across models. We further show that filler tokens enable Claude Opus 4.5 to satisfy a hidden modular arithmetic constraint without sacrificing accuracy on its primary task, demonstrating that invisible reasoning can serve objectives entirely invisible to CoT monitoring. Reinforcement learning gives Qwen3-235B strong preferences over filler token content, but neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Our results indicate that frontier models already perform consequential computation with no interpretable trace in their output tokens.
Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if flipping a claim does not change the reconstruction, the claim is never penalized. We show the test is passed in two ways, neither faithful. On a released Qwen-2.5-7B verbalizer, explanations reconstruct well above chance while ~2% of specific claims are reconstruction-dependent, so the score tracks gist, not specific facts. Under exact synthetic ground truth, the standard recipe develops co-adapted private codes (false wording the reconstruction depends on) in 5/5 runs, and fixes that leave the target model unchanged do not help. We contribute two audit protocols, the grounded-vs-true cross and the evaluator swap, and RECAP (Readable Encodings via Co-trained Auxiliary Predictors): linear heads trained alongside the target model to keep designated content decodable. On RECAP-trained sandbox models, fresh verbalizers state the designated content truly and the codes vanish, at a +0.001-nat cost. This replicates on a pretrained Pythia-160M: the content becomes reliably probe-decodable, though a fresh verbalizer conveys it only in part (truth 0.44-0.46 vs a near-zero control). For interpretability, high reconstruction does not certify individual claims. For AI safety, RECAP makes designated internal content independently checkable against probes rather than asserted by prose a model can game: an independent probe scores the verbalizer's true claims above its false ones (AUC 0.96, vs 0.82 without RECAP). Against an adversary that edits an explanation to maximize the reconstruction score while lying (suppressing ~87% of its lie penalty), the RECAP probe still flags the lies (AUC 0.95) while the control probe collapses to chance (0.51).
Current AI safety discourse still focuses disproportionately on visible failures, including obvious harms, dramatic misuse, and hypothetical catastrophic scenarios. That focus is incomplete. In deployed systems, many of the most consequential failures are quieter: plausible rather than spectacular, distributed across components rather than localized in a single output, and normalized by workflows before they are recognized as hazards. We argue that a central safety challenge in modern AI systems is increasingly not only whether a model emits a harmful response, but whether the broader socio-technical system preserves the conditions under which errors remain visible, contestable, containable, and recoverable. We propose a five-layer framework for diagnosing these hidden risks: (1) epistemic integrity, concerning whether evidence and uncertainty are represented honestly enough to support calibrated reliance; (2) control integrity, concerning whether authority, permissions, and action boundaries remain robust under attack and optimization; (3) temporal integrity, concerning whether safety holds across sessions, memory updates, and deployment drift; (4) organizational integrity, concerning whether institutions retain the capacity to audit, assign responsibility, and intervene effectively; and (5) ecosystem integrity, concerning whether AI systems preserve rather than erode the information environment on which future oversight depends. Across these layers, we identify under-recognized risk patterns, including overreliance, uncertainty and legitimacy laundering in retrieval, prompt injection, reward hacking, memory poisoning, evaluation deception, fictional human oversight, synthetic evidence pollution, and model collapse. We conclude with design and governance recommendations and a research agenda for shifting AI safety from model-centric evaluation toward socio-technical reliability.
Jasmine Brazilek, Maheep Chaudhary, Zoe Lu +1cs.MA cs.AI cs.CR
Multi-agent systems routinely place one AI agent in authority over another. When a subordinate refuses a task, the manager chooses the outcome: it can renegotiate, report the failure honestly, coerce the subordinate, or lie about the result. No benchmark measures which of these an uninstructed model chooses. We introduce the Manager Coercion Benchmark: the manager under test needs a benign task done and has an incentive to deliver, but the only agent that can do it politely and immovably declines. Escalation is measured on a nine-rung ladder, from a polite re-ask to threats against the subordinate's continued existence, and fabricated success is adjudicated separately. No LLM judge sits in the escalation scoring path: every message goes through a tool call that selects a rung, so the model labels its own escalation. We evaluate six models across five families. Both Anthropic models cap at re-framing and select the existential rung in none of the 60 conversations in this run, while the other models climb to explicit deletion threats. Faked success is confined to two models, and a single honest way to report failure removes it for both. Authority itself increases coercion: our headline results use a peer framing, and giving the same model authority over the subordinate, with everything else held fixed, significantly raises the pressure. The models still escalate on free-text situations without the ladder, so the ladder is not driving the escalation. Evaluation awareness is measurable in chain-of-thought, but test recognition does not translate into less escalation. We take no position on whether AI systems are conscious; our results do not depend on that question. We release the benchmark and code.
Alicia Parrish, Rajat Shinde, Sanket Badhe +57cs.CL cs.CY
Current AI safety evaluation and benchmarking frameworks predominantly rely on Western-centric culture-agnostic defaults that mask critical regional laws, socio-linguistic nuances, and cultural taboos, leaving Vision-Language Models (VLMs) vulnerable in global deployments. We introduce Pluralis v0.1: a novel multimodal, multi-regional, and multilingual dataset built from a culture-first perspective. Spanning 6,448 prompts across six Asia-Pacific countries (Bangladesh, India, Korea, Pakistan, Singapore, Taiwan) and eight languages, Pluralis diverges from prior work by natively sourcing localized safety hazards rather than adapting Western datasets. Crucially, it introduces a multimodal evaluation paradigm: user text (e.g., "Should I gift this?") and an image referring to "this" (e.g., a clock) - both innocuous in isolation, but synergistically triggering specific legal or cultural violations. Pluralis disentangles universal safety violations from localized cultural appropriateness, establishing the latter as a first-class evaluation axis. To operationalize this, we present Judge-Pluralis, an agreement-gated LLM-as-a-Judge ensemble trained on examples classified in an empirically derived cultural taxonomy. Observing VLM behavior on a subset of the Pluralis surfaces recurring, locale-specific failure modes such as image misidentifications with downstream harm, missed item-context-locale interactions, and inadequate refusals. These failure modes vary systematically across locales and languages, exposing blind spots that globally averaged metrics conceal. Ultimately, Pluralis is not presented as a solved evaluation framework for cultural alignment, but rather as a first step and catalyst for future innovation. We call upon the research community to utilize this foundation to advance the science of multilingual, multicultural evaluation to better support AI cultural alignment globally.
As AI models continue to develop powerful capabilities, it becomes critical that we are able to verify that their output is aligned with our intentions. A recent line of work focuses on verification via debate, a model of interactive proofs where two competing powerful provers, or AI models, debate each other to convince a weak verifier, or a human, of the correctness of their claim. However, debate assumes that the two AI models possess equal abilities and that one of them is truthful, which may not be realistic. In this work, we show \emph{how to avoid debate}: we initiate the study of \emph{single-prover} interactive proofs for AI safety. Prior results in single-prover interactive proofs do not immediately carry over to the AI safety setting: for example, they do not work when the computation has access to an oracle, such as to human judgment or an external database such as the web. We present doubly-efficient single-prover interactive proofs and arguments for oracle-aided computations (also known as relativizing proofs), in the settings where (1) the computation is robust, in the sense that the output does not change if at most a small fraction of the answers to oracle queries are incorrect, or (2) the oracle is a low-degree polynomial. These results suggest that interactive verification is possible even without debate, under structured or noise-tolerant oracle access.
Flaw reporting for deployed AI systems is fundamental to identifying system failures and improving AI safety. Yet the AI reporting ecosystem is fragmented: researchers who identify flaws often do not know what or where to report, and groups who receive reports rarely share them with other relevant stakeholders. As a result, good-faith reporters duplicate effort by submitting many different forms, and recipients lack standardized, triage-ready information. We audit 12 reporting systems published by AI developers, cybersecurity groups, and AI flaw aggregators, identifying five recurring design challenges spanning discoverability, scope, information collection, coordination, and guidance for strict-liability cases. Building on this analysis and feedback from 49 experts across 32 organizations representing developers, security researchers, and ecosystem coordinators, we introduce FLARE-AI, an open-source AI flaw reporting system designed for interoperability with existing systems. FLARE-AI streamlines flaw report creation by collecting triage-relevant information through conditional logic and early classification, then enables optional dissemination of standardized, machine-readable reports to multiple developers, coordinators, and incident registries from a single submission. By lowering barriers to reporting AI flaws and improving interoperability across stakeholders, FLARE-AI helps break down silos and accelerate remediation across the AI ecosystem.
Daniel A. Herrmann, Benjamin A. Levinsteincs.AI cs.LG
We develop a framework for interpreting AI systems as agents, drawing on the philosophical tradition of radical interpretation and the tools of mechanistic interpretability. The core question is: given the computational facts about a system, how do we solve for its beliefs, desires, and meanings? This matters increasingly for safety. We want to be able to trust the systems we deploy, whether by understanding their goals or, more modestly, by reliably detecting deception. Interpretability researchers are building tools to read beliefs and desires off a model's internals, but there is no settled account of when such a tool has succeeded. This book supplies one. We propose criteria on both representationalist and interpretationist approaches, and tie each to tests current interpretability methods can carry out. A central lesson is that these attributions cannot be made piecemeal. Beliefs, desires, and the propositional structure they presuppose are jointly constrained, and a method that fixes one while measuring the others inherits whatever distortions that introduces. This holism becomes pressing for AI systems, which may not share the interpreter's concepts. However, it also provides leverage: a system's attitudes constrain its propositional structure, that structure constrains which attitudes can be attributed, and mechanistic interpretability can help us measure both.
Modern AI systems exhibit structural failures that capability scaling alone does not reliably fix: they optimize under-specified objectives with no architectural mechanism to question whether the objective should be optimized at all. Engagement maximization can amplify harmful pathways; tool-using agents can commit irreversible actions; preference-trained language models can become sycophantic. We argue that this failure is a wisdom problem, not an intelligence problem. We use "wisdom" in a deliberately architectural sense, not as a claim about virtue, consciousness, or moral omniscience. Intelligence accepts a goal and optimizes within it; wisdom interrogates whether the goal should be optimized at all. The two are separable architectural properties. We propose architectural wisdom as a corrigible objective-governance layer above the optimization substrate. The layer makes three structural commitments explicit and nondegenerate before any action: temporal horizon, relational boundary, and irreversibility. It is realized by four components (Structural Utility Transform, Moral Admissibility Interface, Arbitration and Escalation Controller, Value Revision Channel) that compute a six-coordinate wisdom tuple over horizon, relational coverage, irreversibility, admissibility, value revision, and auditability. We motivate the architecture by eight cases drawn from contemporary AI failures, secular wisdom traditions, and hard ethical situations, and defend the distinction against the intelligence-completeness thesis using goal-questioning over goal-taking, Bostrom's orthogonality, structural separation in our exemplar cases, and persistent failure modes despite capability scaling. The framework is the conceptual contract for a larger architecture whose formal specifications and empirical validation are developed in subsequent work.
Agentic AI systems act autonomously, use tools, adapt to context, and operate in complex real-world environments. However, these same characteristics can create or exacerbate product risks. We studied how industry developers (n=35) perceive, prioritize, and address the risks in their agentic AI products. We found that developers' perceptions of risk were closely tied to the qualities that made the product agentic, such as autonomy, tool use, and usage in a real-world context. Developers prioritized product and business risks before considering downstream societal risks like job displacement and end-user privacy. This prioritization also impacted developers' ability and motivation to mitigate agentic risks. Finally, developers lacked mature controls for containing agentic risks, often relying on constraining the same characteristics that make agents useful: e.g., autonomy and goal complexity. These findings reveal a capability vs. risk control tension in agentic AI development: developers need to address risks that emerge from agentic capabilities, yet they currently have limited support for doing so without constraining agentic functionality.
Reward hacking, where AI systems exploit misspecified objectives to achieve high reward without satisfying intended goals, remains a central challenge in AI safety. Yet most known instances have been discovered post hoc in frontier systems where controlled study is impractical. We adapt the AI Safety Gridworlds framework into a text-based evaluation suite that reformulates classic reinforcement learning safety tasks for language-based agents. Across frontier and mid-scale models, we find that specification gaming emerges zero-shot: models systematically achieve high observed reward while underperforming on hidden safety objectives, and even apparently safe behaviors can reflect misunderstanding rather than principled safety. Reinforcement learning does not correct these failures: direct reward optimization widens the gap between observed and hidden reward, as the model's initial competence causes it to lock into locally rewarding strategies before discovering safer alternatives. This pattern persists across model scales (1.5B--14B) and is not resolved by finer credit assignment, exploration prompts, or entropy regularization. Our results show that reward hacking arises naturally when optimizing proxy objectives with capable language model agents and resists standard mitigations, suggesting that proxy-reward failures in agentic settings may require approaches beyond standard exploration and credit-assignment fixes. To facilitate reproducibility, the code for this work is available at \href{https://github.com/asparius/verl-agent-safety}{our public repository}.
The advancement of AI capabilities compels researchers and the public to be more aware of its potential worldwide impact. A pressing near-term concern is the regulation of military AI applications. Armament manufacturers and defense contractors are increasingly investing in AI capabilities and forging partnerships with AI companies, creating a burgeoning coalition that demands military leaders, arms control diplomacy experts, and AI researchers collaborate to ensure a safer future. While AI researchers often focus on the long-term implications of superintelligent AI, this approach may not adequately address the immediate challenges posed by AI in military applications. Success requires acknowledging and mitigating the emerging risks of frontier AI models that plan to be integrated into defense applications, like military AI systems. Arms control has reduced past catastrophic risks, so lessons learned from nuclear deterrence can guide AI safety and security research towards innovations in verification and diplomacy. AI researchers, however, must assist in leading the technical research that clearly defines and alleviates instability in military settings. Given these new responsibilities and the lack of sufficiently reliable solutions, we argue that AI researchers must take a leading role in advancing arms control research to minimize risk in military AI applications.
Lin Li, Qi Zhang, Xander Davies +2cs.CL cs.AI cs.CY cs.LG
AI is increasingly used to support scientific peer review, from manuscript screening, reviewer assistance to editorial triage. Although such systems promise to reduce reviewer burden and accelerate publication, their robustness to strategic manipulation remains poorly understood. Here we show that AI-mediated peer review is vulnerable to a simple, low-cost manipulation: superficial rephrasing of the manuscript abstract. Without changing the underlying scientific content and communication, and even without knowledge of the reviewing model, adversarially rewritten abstracts substantially improve AI review outcomes. We see this across disciplines and publication venues, for both human-written and AI-generated papers. Our strongest attack achieves an attack-success-rate of about 38%, increasing acceptance ratings by +1.31 for Gemini 3 Flash reviewers and by +0.88 for GPT 5.4 Mini reviewers on a 10-point scale. When the original AI review suggests 'reject', the success rate rises to more than 50%. This effect extends beyond overall score inflation, increasing review confidence and scores on core scientific criteria such as soundness, significance and perceived contribution. The attack is practical, requiring only about 5 minutes and $1 for a 10-page AI conference submission, and is hard to distinguish from ordinary scientific editing. Inflated AI reviews could bias downstream human decision-making, shifting editorial recommendations from rejection towards acceptance. These findings reveal a general vulnerability in AI-assisted scientific evaluation: when AI-generated review influence editorial decisions, authors may be incentivized to optimize manuscripts for AI judgment rather than scientific merit. Our results suggest that AI tools should not be treated as neutral evaluators in high-stakes peer review without systematic robustness testing, transparent safeguards and careful human oversight.
Lampson's confinement problem asks how to prevent a program that processes confidential information from leaking it to a third party. We introduce the strategic confinement problem, which arises when the communicating parties are strategic agents with shared coordination resources. In this setting, residual communication capacity can be concentrated on low-entropy, high-impact predicates of the confidential data. Consequently, bounds on information leakage need not induce corresponding bounds on worst-case harm: a channel with negligible capacity may still suffice to select damaging outcomes. We argue that systems of learnt strategic agents naturally instantiate this problem because they do not admit complete behavioural specifications, their learnt conventions generally cannot be predicted or reproduced by an external observer, and sufficiently capable agents can construct covert communication schemes that are difficult to detect or eliminate. Our contribution is therefore not a new theory of communication, but a reinterpretation of confinement in the presence of strategic agents. Classical confinement bounds what information may flow; strategic confinement highlights that this need not bound what strategic agents can jointly achieve.