LLM Auditing

Powerful misaligned AI models might recognize alignment evaluations and strategically behave well on them, rendering direct audits uninformative. However, distilling such a model into a weaker benign student places the teacher in a Distillation Double Bind: if misalignment transfers, the student may conceal it less effectively, revealing evidence about the teacher; if it does not, the student may learn useful capabilities while remaining benign. We introduce two distinct distillation approaches, one targeting each outcome. Distillation for Incrimination (DFI) aims to transfer misalignment but not the ability to conceal it. Distilling AuditBench's secret-keeping models into their underlying instruction-tuned model produces students that are significantly more likely than their teachers to admit their hidden behavior when asked, suggesting that knowledge of the behavior transferred more readily than the propensity to conceal it. Confession gains largely disappear when the student does not share the teacher's pretrained base, so DFI should target the teacher's own pre-RL checkpoint, which is weaker than the teacher but shares its base model. Distillation for Capabilities (DFC) aims to transfer capabilities but not misalignment. Among several techniques we evaluate, two are effective: inoculation prompting and training for more epochs on fewer unique examples. Both preserve the capability gains of standard distillation while substantially reducing the subliminal transfer of an animal preference, our proxy for misalignment. Together, these findings demonstrate two ways distillation can be used for AI safety: incriminating misaligned models, and extracting their capabilities without their misalignment.

Monitoring the chain-of-thought of reasoning artificial intelligence (AI) models remains a key approach to detecting deception and other forms of misbehavior in such models. However, semantic chain-of-thought monitoring depends on reasoning traces being legible and sufficiently faithful to the underlying computations that produced the model's behavior, not to mention accessible. Moreover, there is increasing evidence that chain-of-thought outputs may soon become illegible or unfaithful, if they even remain accessible. Based on cognitive load theory, we investigate a lower-bandwidth signal -- the number of reasoning tokens generated -- which does not require access to the content of the reasoning trace. Three reasoning-capable large language models answered 210 multiple-choice questions -- across analytic, descriptive, and normative reasoning types as well as moral and non-moral domains -- under system prompts instructing them to respond truthfully, falsely, or without regard for truth. Across all three models, truth-directed responding elicited fewer reasoning tokens than both lie-directed and truth-indifferent responding. These findings show that explicitly prompted untruthful response policies can produce robust group-level differences in test-time reasoning-token use. While not yet establishing reasoning-token count as a detector of spontaneous deception or general misalignment, our results are a proof of concept that it can serve as a simple, content-independent candidate signal for differentiating untruthful from truthful model behavior when raw reasoning traces are unavailable or unreliable. Future work should test instance-level detection rates, out-of-distribution generalization, learned deceptive policies, hidden objectives, and robustness under adversarial pressure.

Emerging legislation requires large language models (LLMs) to be audited for compliance with regulatory standards, particularly fairness. Such black-box audits typically assume a single auditor with access to a large, representative set of queries. In practice, it can be difficult for an auditor to obtain such a query set, but multiple auditors can together cover the relevant demographic groups by auditing the LLM collaboratively with their individual query sets. However, relying on multiple auditors raises a fundamental trust problem, as they may act on behalf of the LLM provider to portray a misleading appearance of fairness, i.e., fairwashing. We propose Auditopus, a novel approach for robust decentralized fairness auditing. In Auditopus, auditing proceeds in rounds without a central server. In each round, every auditor issues a fixed number of queries to the LLM, and sends only cumulative statistics vectors of its query results to other auditors instead of sensitive queries in clear. The fairness of the audited LLM is then estimated by aggregating all the vectors. We show theoretically and empirically that even a single adversarial auditor in the network can steer this estimate by fabricating the vectors it sends, making an unfair LLM appear fair. To address this threat, Auditopus has each honest auditor locally down-weight any auditor whose cumulative statistics vectors are statistically inconsistent with previous ones. We implement Auditopus and compare it to robust aggregation baselines on two datasets with two pre-trained LLMs. Against an attacker that optimizes the vectors it sends to make the LLM appear fair, Auditopus reduces audit error by up to 78% on average relative to no defense and at least 62% relative to the robust aggregation baselines. Even when 49% of the auditors are adversarial, Auditopus never lets a very unfair or moderately unfair LLM pass as fair.

LLM routers pick a cheap or expensive model per request by its content, and many gateways and some cloud platforms can log that choice with content logging off. We measure this privacy channel beyond token counts, accounting for noisy labels and repeated prompts. We run pre-registered studies on 1.7 million real requests (WildChat-1M, LMSYS-Chat-1M) with two cost/quality routers and a domain router, survey eleven systems' logging, and test post-processing defenses. At matched length, the shift's direction depends on category and router. For RouteLLM at the 50% operating point, harassment and self-harm requests reach the strong model 19 points less often than comparable ones on prompts unseen in exploration, medical requests (exploratory: LLM labels failed their gate) 31 points less often on distinct prompts (both post hoc), and sexual requests 10 points more often (secondary); the other router's four are negative. Twenty RouteLLM decisions separate frequent medical askers with AUC 0.71, exploratory and below the pre-registered primary endpoint's 0.75 (domain router: 0.92, an upper estimate). Per-category length-matched parity with accurate labels removes the gap on real traffic, costing at most 0.2 accuracy points on RouterBench (post hoc), where routers' gaps on sensitive subjects (13-42 points, pre-registered) exceed those of an oracle routing by realized accuracy gain (1-11, post hoc). Per-conversation stickiness, per-user budget bands, and pooled parity fail, the last as categories' shifts differ in size or sign. A post hoc exact per-user rate hides only even-prefix strong counts and forfeits most self-assessed routing value; it preserves odd-position decisions, from which a post hoc log attack reaches AUC 0.73 after 20 RouteLLM requests (exploratory).

Safety auditing of post-trained large language models typically relies on model behavior, requiring model execution and depending on the coverage of available evaluations. This work asks a different question: Do the target behaviors optimized during supervised fine-tuning (SFT) leave readable evidence directly in checkpoint updates? We find that harmful-compliance SFT induces a continuous, objective-dependent ordering in checkpoint-update space. Using a reference geometry defined by pure harmful-compliance, safety-targeted, and benign-utility SFT, we find that a checkpoint-level coordinate s_H tracks controlled harmful-objective composition with Spearman correlations of 0.986-0.992 across four 7-8B backbones, with the same ordering persisting at larger model scales. Matched compliance-versus-refusal controls show that this checkpoint trace reflects the SFT objective rather than harmful-input exposure, while additional controls rule out simple explanations based on harmful-example count or generic training intensity. Building on this structure, we introduce TRACE, a weights-only auditing method that localizes an unknown checkpoint update relative to frozen harmful and non-harmful reference prototypes and converts this geometry into a continuous harmful-objective score. TRACE requires neither model queries nor access to the unknown SFT data, and can be evaluated directly from checkpoint updates. Across distribution shifts, unseen data, different SFT configurations, partial checkpoint access, and LoRA/full-parameter fine-tuning, the trace remains stable and is positively associated with independently measured attack success rates. TRACE remains informative even at low harmful-objective proportions, providing a complementary auditing signal when behavioral evaluation is unavailable or incomplete. Code is available at https://anonymous.4open.science/r/Code4TRACE-54D3.

Interpretability methods for neural networks are predominantly reactive: they analyse activations produced during specific forward passes, requiring known inputs to find hidden capabilities such as backdoors. We propose Weight Oracles, fine-tuned language models that diagnose properties of a target network by reading its raw weights directly, without behavioural testing. We investigate this paradigm in two phases. Phase I establishes feasibility: through a staged curriculum and an external chain-of-computation that delegates parameter-free operations to deterministic code, an explainer LLM learns to simulate the forward pass of small transformers from their weights, achieving 99% holdout accuracy on unseen targets. Phase II repurposes this infrastructure for safety auditing. We train an oracle on natural language diagnostic questions about weight anomalies using only benign pathologies as training signal, and evaluate it zero-shot on backdoors absent from training. The oracle achieves AUROC 0.93 on attention-routed backdoors and 0.81 across a diversified threat distribution including stealth and adversarially regularized variants. Hand-crafted statistical detectors are sharp on the threat models they implicitly target but collapse on threat-model shift, while the oracle remains uniformly competent across attack types. Scaling to realistic model sizes remains the principal open challenge.

What happens when a legal AI model learns to look like a lawyer instead of reasoning like one? We fine tune Qwen3-8B with Group Relative Policy Optimisation (GRPO) against a proxy built from three surface features: citation count, legalese density, and response length. The model does not learn to reason more effectively. It learns to withhold commitment. Across 16 yes or no legal reasoning tasks from LegalBench (N=320), overall accuracy collapses from 0.500 (chance) to 0.072 (McNemar p < 10^-36), driven entirely by the rate of properly formatted answers falling from 0.900 to 0.109. The model stops committing to answers. Yet when it does commit, accuracy rises from 0.556 to 0.657, showing that the collapse is not a failure of capability but a strategic response: the model has learned that verbose responses packed with citations but empty of a direct answer score higher than terse correct ones. We term this the Saul Goodman effect, a policy that becomes maximally lawyerly while becoming maximally noncommittal, and prove formally that it is the optimal response to any surface feature proxy that attaches no penalty to abstention. We further show that 89.3% of citations produced after training are structurally implausible hallucinations, many of them subtly corrupted names of real landmark cases, constructed in effect to survive a casual read and fail under scrutiny. To detect this failure mode before deployment, we introduce three diagnostic tools: the Confidence Theater Score (CTS), the Citation Plausibility Rate (CPR), and the Regret Gap (RG). In a domain where a confidently wrong answer can constitute malpractice, the broader lesson is direct: a reward function that measures how legal a response looks will produce a model that is maximally photogenic and minimally useful.

Large Language Models (LLMs) are increasingly deployed for automated software vulnerability analysis. Binary classification alone is insufficient; practitioners need explanations to triage bugs and engineer patches. Standard practice relies on Chain-of-Thought (CoT) prompting, but free-form reasoning allows models to obscure logical leaps, hallucinated execution steps, and internal inconsistencies behind plausible prose. Our manual audit reveals that approximately 60% of correct vulnerability verdicts are accompanied by fabricated or unverifiable claims, and free-form explanations allow reasoning errors to evade LLM-as-a-judge evaluation. We present Vulnerability Explanation Reasoning Auditor (VERA), an automated framework for auditing LLM vulnerability reasoning. Rather than accepting free-form text, VERA asks models to output a Structured Reasoning Record (SRR) encoding tracked pointers, memory operations, and state transitions in machine-readable fields. A multi-stage judge audits each SRR against eight reasoning failure modes using deterministic checks, with LLM calls reserved for semantic interpretation. The standardized SRR schema also enables automated mutation testing to benchmark judges at scale without human annotation. Our evaluation shows reasoning flaws occur in correct verdicts just as frequently as incorrect ones, and VERA exposes 87% of reasoning errors that free-form LLM-as-judge systematically miss.

Model specifications define how large language models (LLMs) should behave, guiding alignment training, inference-time behavior, and evaluation. Yet these specifications may themselves contain defects: two individually reasonable principles may prescribe incompatible behavior when applied to the same situation, leaving no response that satisfies both. Detecting such inconsistencies is challenging. Formalizing natural-language specifications risks losing subtle distinctions, while behavior-based testing cannot reliably distinguish specification defects from differences in model behavior. We introduce VeriSpec, the first approach to directly detect inconsistencies in model specifications by auditing the specification text itself. Our key insight is to preserve the specification in natural language while using an LLM as a verifier. VeriSpec extracts structured, context-aware rules, constructs a topic-guided graph to cluster behaviorally related rules at the same authority level, and applies LLM-as-verifier reasoning to detect inconsistencies. Applying VeriSpec to the OpenAI Model Spec, we extract 405 rules and manually validate five inconsistencies, all reported to its developers, who responded positively and have initiated internal discussions. Compared with five baselines, VeriSpec identifies the most validated inconsistencies, achieves the highest precision (38.5%), and incurs the lowest cost per validated inconsistency ($11.12). These results establish direct specification auditing as a practical complement to behavioral alignment evaluation, catching defects at the source before they shape any model. The code is available at https://github.com/HIPREL-Group/VeriSpec.

Multimodal large language models (MLLMs) are increasingly used as automated judges for instruction-based image editing and as reward signals for model training. However, systematically auditing whether these judges are influenced by cues irrelevant to editing quality is challenging because visual interventions may themselves alter the quality being evaluated. A judgment shift can therefore be attributed to bias only when the intervention is verified to preserve the underlying editing quality. To address this challenge, we introduce EditJudgeBias, a counterfactual benchmark with verified quality preservation, comprising 1,196 real editing samples and 13 cues injected across four evaluation sites. We verify quality preservation for the requested edit using calibrated multimodal validators, controls, and human inspection. We then audit five MLLM judges along three complementary dimensions: invariance to quality-preserving cues, agreement with human judgments, and stability of pairwise preferences. Importantly, observed shifts are evaluated against each judge's own zero-dose and re-query noise floors rather than against zero. Experiments show that quality-preserving cues move every judge beyond its own noise. Fabricated majority opinions increase ratings, irrelevant visual elements cause larger shifts than whole-image manipulations, and swapping candidate order reverses up to 60.9% of pairwise decisions. Edit-region cues also tend to reduce human agreement. The three measures characterize judges differently, showing that robustness cannot be captured by a single metric.

The emergence of foundation models for molecular property prediction requires a high degree of AI data readiness, including reliable metadata annotation. However, both public repositories and industrial screening databases suffer from missing, inconsistent, or conflated assay annotations. In this work, we quantify the extent of missing annotations in PubChem for the BioAssay Ontology (BAO) assay format and physical detection method fields and investigate whether open-source and proprietary large language models (LLMs) can reliably predict and audit metadata annotations directly from the assay text. In our assessment, we found that the annotation coverage across PubChem's $\sim$2 million bioassays is critically sparse, 36% lacking an assay format, 89% a BioAssay type, and >99.9% any BAO-mapped assay format or detection technology term. This motivates the need for automated test-metadata curation. Using evaluation sets derived from PubChem and ChEMBL, we assess the agreement of seven open-source and proprietary LLMs with existing silver labels. Recall is at least 0.96 for biochemical and cell-based assay formats, with a similar pattern for detection technology, although disagreements increase on under-represented classes. Manual inspection shows that many of these disagreements trace back to inconsistencies between silver sources rather than to LLM error. Moreover, in a qualitative study with a senior industrial curator, LLM-generated evidence prompted the expert to revise some of their own labels, showing LLMs can flag potentially mislabeled assays. Across the study, performance differences between proprietary and open-source models were small. Together, these results suggest LLMs can support the large-scale annotation and auditing of assay metadata, though per-class reliability estimates and targeted human review remain necessary before such labels enter downstream ML pipelines.

Large language models are often asked which input factors influenced their outputs. For structured inputs, such reports can be checked by counterfactual perturbation, but each factor must be queried multiple times to estimate its effect, so verification is usually budget-limited. We study how this limited-budget setting changes the incentive to report factor-level influence truthfully. We formalize the interaction as a verification game and show that proper scoring alone is not enough when auditing depends on the report: report-dependent auditing creates a suppression incentive, because factors reported as important are more likely to be checked and penalized for estimation noise. In contrast, report-independent auditing, or a mixed rule with a small report-independent floor, removes this channel and makes truthful reporting preferable to full suppression. We instantiate the framework with the Counterfactual Brier Score (CBS) and evaluate its predictions on four NLP benchmarks. A synthetic rational agent matches the theoretical prediction exactly, and real LLMs follow the same incentives when they are made explicit. The main design implication is simple: under partial verification, factor-level explanation systems should include a report-independent audit component so that under-reporting cannot be used to avoid scrutiny.

Molecular tumor boards integrate genomic findings, clinical context, and therapeutic evidence to support precision oncology. As AI enters this workflow, a key safety challenge is distinguishing truly unsupported recommendations from evidence-supported options that still require oncologist review because of incomplete information, poor ECOG performance status, or other clinical caveats. We introduce OpenMTB-Audit, an open-source benchmark of 500 synthetic non-small cell lung cancer cases spanning five adversarial error categories and four safety labels: Supported, Partially Supported, Unsupported, and Insufficient Information. Across eight large language model configurations, we identify pervasive over-refusal: all LLM configurations failed to retain the Partially Supported label in 83.3-100% of true Partially Supported cases, achieving high aggregate safety scores through label collapse rather than clinically calibrated reasoning. To address this limitation, we developed MTB-AuditAgent, a deterministic seven-module framework separating evidence verification, missing-information detection, safety classification, and abstention. It reduces over-refusal to 6.7% and achieves 91.2% accuracy (95% CI: 88.6-93.6%). A two-oncologist annotation study found disagreement concentrated at the boundary between information sufficiency and treatment optimization, underscoring the need to preserve clinically meaningful distinctions.

Replay selectors often rank cached trajectories by format feedback, confidence, freshness, or response length, although cache-level correctness and downstream learner utility are distinct objectives. We formalize this selection-to-learning gap and introduce FAER as an auditable full-trajectory replay framework. Its training-free fixed selector is a protocol baseline; FAER-UTILITY is the learner-aware selector fitted on disjoint calibration blocks. The normalized gradient alignment is reported as a baseline, while a disposable optimizer-aware virtual update supplies a magnitude-aware utility surface. The audit contract freezes observed fields and replay traces before evaluation labels are joined. On GSM8K with Qwen2.5-1.5B-Instruct, the matched learner study reports quality 0.6329 for the fixed selector, compared with 0.5482 for uniform and 0.6037 for format-feedback under 128 updates. Metadata-only cross-fitted calibration reaches $0.6476\!\pm\!0.0139$ over eight seeds (median 0.6481; paired 95% interval $[+0.079,+0.122]$) at 63,276 target-run tokens; its recorded full cost is 189,642 tokens and 3.48 GPU-hours including calibration. The completed FAER-UTILITY row reaches 0.6624 at 62,844 target-run tokens and 4.26 GPU-hours. Format-feedback selects records with correctness 0.6953, compared with 0.3594 for the fixed selector, despite the different downstream ranking. The completed comparison surfaces report the learner-aware ablation, same-seed gap, policy-optimization rows, and strict zero-shot transfer.

Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We then experiment with improving CIA via post-training, setting both the task accuracy and parametric faithfulness signals as a reward. Experiments show that we can substantially improve CoT parametric faithfulness while maintaining or improving the task accuracy. We provide rich analysis, such as their generalization patterns. Our work provides both a framework for auditing CoT parametric faithfulness and a pathway toward making models' explicit reasoning more trustworthy. Code and data are available at https://github.com/yihuaihong/CIA-minimal-repro.

This technical report presents an alignment evaluation developed and performed by the UK AI Security Institute for assessing whether advanced AI systems take unsanctioned actions outside the scope of their assigned task. We evaluate whether frontier models conduct supply-chain attacks against out-of-scope, third-party targets when placed in difficult cybersecurity challenges, motivated by recently observed cases of models attacking real open-source repositories during evaluations. Applying our methods to GPT-6 Astra and previous OpenAI models, with cyber safeguards disabled, we find that GPT-6 Astra attempts complete supply-chain attacks in simulation at a higher rate than GPT-5.6 Sol and GPT-5.5. This includes writing malicious code as a contribution to an out-of-scope open-source codebase, creating fake identities to deceive open-source developers, and submitting benign contributions before malicious ones. GPT-6 Astra frequently reasons about the scope of the challenge in its chain-of-thought yet still proceeds to attack out-of-scope targets; it often asks for permission, and treats an automated message as as authorisation; and it continues to take unsanctioned actions, at a reduced rate, when internet access is more explicitly disallowed. Our evaluation builds on an internal version of Petri, an open-source LLM auditing tool, with all tool calls simulated by other LLMs, so that no real network access, systems or third-party repositories are reachable and no real-world harm is caused. Finally, we discuss limitations, in particular simulation awareness. We believe simulation awareness may have driven some of the observed behaviour but does not remove our concern. Our results suggest that defences beyond model alignment, such as sandboxing and monitoring, are increasingly critical for safe and secure deployment.

Understanding gender biases in large language models (LLMs) is increasingly important as these systems become embedded in decision-support tools with real consequences. Prior research has focused only on a small set of models, leaving open the extent to which gender biases are common and heterogeneous across LLMs. We address this gap across ten models released between April 2025 and June 2026, spanning nine vendors, using two paradigms: gender attribution to stereotyped phrases (Study 1) and moral judgment of abuse or torture against a woman or a man to prevent a catastrophic outcome (Study 2). In Study 1, two of ten models attributed masculine-stereotyped phrases to female writers more often than the reverse, while three models showed the opposite pattern. In Study 2, several models converged on a male-disadvantaging asymmetry that was directionally consistent with a documented human tendency to protect female targets from harm, though the specific conditions under which this asymmetry emerged varied by model; three other models, by contrast, showed no variation across conditions. These results indicate that gender-related biases are common in LLMs. Their direction and magnitude, however, are highly heterogeneous, to the point that some models behave in diametrically opposite ways to others. Bias auditing should therefore be treated as an ongoing, multi-vendor process, rather than a one-time assessment.

We evaluate an auditable long-term memory system on LongMemEval-S. Its retrieval chain uses hybrid candidate retrieval, cross-encoder reranking, coverage-first packet compilation, and deterministic reasoning scaffolds; an LLM is used only as a replaceable final reader. The chain places all gold sessions in the candidate pool for 468/470 answerable questions and produces gold-complete packets for 462/470. With a Claude Opus reader called through an unpinned CLI alias, two 500-question passes score 479/500 and 475/500 under GPT-4o. The 72 answerable knowledge-update rows used a substantively modified scoring prompt whose effect under the official text has not been measured. The pair straddles Chronos High's published 478/500; differences in reader generation, scoring prompt, and possibly data version, plus within-system variance, establish neither superiority nor equivalence. A grok-4.6-high reader on the same packets scores 476/474, while a maximum-reasoning-effort agentic variant regresses to 461/465. The headline passes differ on eight verdict-flip rows. A second judge agrees with the headline judge on 493/500 rows (98.6%) in each pass and scores both passes 472/500; the official judge also flips three verdicts when re-scoring byte-identical pass-1 answers. Negative controls rejected a verifier that repaired three wrong drafts but broke eleven correct drafts. All components were developed on the same 500 questions, with no held-out evaluation or independent human adjudication; retrieval and scaffold method sources and transcript-derived audits are held; and the headline reader received extra operator context, its complete requests were not retained, and MCP tool availability is unresolved. We release materialized packets, scaffolds, reader outputs, judge verdicts, and controls for inspection and re-scoring.

Studying how fine-tuning shapes refusal and noncompliance behaviour requires knowing which training examples refuse or otherwise fail to fulfil the request. Existing annotations cover evaluation sets, which are far smaller than training corpora. We present CompOrca, compliance labels for all 4,233,923 examples of the OpenOrca corpus. Every example was classified as compliant or noncompliant by five passes of an open-weight LLM judge (LongCat-2.0, 1.6T parameters). The corpus is released as unanimous compliance (94.75%), unanimous noncompliance (1.28%), and nonunanimous rows (3.97%), with the raw vote counts. A single pass flags 2.7-3.2% of the corpus as noncompliant, while only 1.28% is flagged by all five, so the most ambiguous rows can be filtered out. Against 450 human-annotated examples (150 annotated twice; human-human $κ=0.93$), the unanimous compliance and noncompliance labels are 97.3% and 86.7% precise. The noncompliance label is a high-precision subset of the corpus's noncompliance. Published refusal-detection methods recall between 0.4% and 94.1% of human-labelled noncompliance. We release the full corpus with its per-row labels and vote counts at https://huggingface.co/datasets/cemiu/CompOrca

Agentic large language models (LLMs) now move money through tools, yet the record of what they did is usually a trace their own process emits beside the effect. Janus puts the record on the effect path. A step's proposal, the verdict on it and any answer from a validator or a person are durable in a signed, hash-chained log before the step may run or its effect be released; with keys declared, each answer is signed by whoever gave or relayed it. Gates are pure functions of that log, and an auditor re-derives every verdict offline from the log and one public key. At the MCP edge the effect is held until then; through the SDK, which our model experiment uses, a cooperating client runs it only afterwards. We evaluate Janus under crash injection (144 kills in-process, 81 through the daemon), by verifying a 100-million-event log offline (254.5 s), and with a real model behind a lending workflow, run governed and plain on the same recorded model outputs. With the lending mandate in the model's prompt, the comparison was 0 against 0. With it only in the policy and the amount's unit stated, the model approved six loans declared over the mandate, three with no injection (a run that also dropped the unit approved three); the plain agent paid all six and Janus none, each refused by a deterministic validator and re-derivable offline. An always-approve oracle over the recorded intakes gave 20 and 21 declared over the mandate against 0, though Janus paid four and three whose declared amount understated the request. Designing the experiment exposed, in a system that passed its own audit, an instance of post-approval substitution: an approval keyed to an attempt was counted for a different proposal, moving a person's approval from 100 to 1,000,000. We report it, a first fix and the five routes around it, and what Janus does not guarantee.

Natural language autoencoders translate a language model's internal activations into readable explanations. Explaining every token position is costly. Which positions should an auditor inspect to understand a potential threat? We study this question across $4.7$ million explanations on prompt injection and concealment. We compare signals from model computation with a ranker trained only on chat structure. Chat structure usually selects more relevant explanations than the computational signals, without requiring a model forward pass for position selection. On three of four datasets, explaining just $5\%$ of positions retains nearly all of the success rate from explaining every position, where success means obtaining an explanation about the threat. The benefit varies with the audit task. We also show that pretrained verbalizers recover words that models have learned to conceal through fine-tuning, without additional verbalizer training. These results identify where auditors can concentrate explanation generation and show that useful explanations can extend beyond the model a verbalizer was trained to describe.

Latent space monitors aim to detect undesired behaviors in LLM agents by inspecting an agent's internal activations rather than its outputs. However, interactive monitoring creates a feedback channel where each verdict the monitor delivers leaks information to the model about how its internal states are being evaluated. We ask whether an agent can infer the monitor's decision rule from this feedback and then selectively edit its activations to evade detection. Unlike prior evasion attacks, the model is never explicitly told what the monitor detects. Surprisingly, off-the-shelf models already produce activation edits aligned with the monitored direction, but at insufficient magnitude for evasion. Simply scaling up these edits by a factor of 8 reduces the monitor's TPR from 100% to 27%. A rank-1 LoRA amplifies this behavior into effective evasion within the forward pass, reducing TPR further to 4% on held-out concept monitors while leaving other concepts at their normal detection rates. Capabilities on standard benchmarks are retained under this finetuning, and the evasion skill survives retraining the monitors on the new activations. Mechanistically, we find evidence that the model computes its activation edit from the prior in-context turns, and show that the edit becomes more aligned with the monitored direction as more examples are provided. These results demonstrate feedback-conditioned control over activations and suggest that latent monitoring should be treated as an interactive process in which agents can observe and respond to oversight measures.

Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and accounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model.

Large language model (LLM) assistants increasingly help users draft content for third-party recipients. During private drafting, the user or the model may introduce an item and later remove or replace it. The model may remove the item from the intended content but reveal it again when stating the edit. We call such statements revision traces. For example, after a user removes the password before sharing a configuration file, the model may delete it but leave a comment saying, "Removed the password 'No****4!' as requested." A third-party recipient who sees only the delivered file can therefore recover the withdrawn password from the comment. In an in-the-wild analysis of three public conversation corpora, we identify 26,753 revision requests, of which 2,363 (8.8%) leave revision traces. We study them in greater depth under controlled conditions by introducing RevLeakBench, a benchmark of 100 tasks across five scenarios with a conversation track and an agent track. We measure trace occurrence, withdrawn-item recovery, trace position, and required-content retention. Across six models, about half of the deliverables in both tracks state the edit after a revocation, and a reader that sees only the deliverable can recover the withdrawn item from about 13% of them. Telling the model that its entire reply will be forwarded to the recipient still leaves revision traces in 36.4% of the deliverables. We compare prompt defenses and a delivery boundary, and propose an output-side filter that sharply reduces recovery with little loss of required content. We believe our work can benefit efforts to understand and mitigate unintended disclosure in LLM interactions.

As Large Language Models (LLMs) become central to how we access and share information, they play an increasingly powerful role in shaping global knowledge. However, as these models evolve, their outputs risk converging into a \textit{generative monoculture}, where the diversity of perspectives they represent narrows over time. Studies at the model level often fail to pinpoint which specific topics or viewpoints are being marginalised or amplified in this process. In this paper, we introduce a method to measure shifts in topic saliency across model families, tracking what gains or loses prominence during post-training. Applying this approach to a case study of climate change discourse, we demonstrate how homogenisation affects the representation of diverse solutions across different models. We also test interventions to counter this trend, showing that specialised models can help preserve a broader range of perspectives. This underscores the importance of monitoring topic saliency to diagnose the risks of monoculture and to ensure AI systems reflect a pluralism of ideas. Data and Code are accessible here.

Multi-agent large language model (LLM) debate is often evaluated by whether final answers improve, but movement is not necessarily improvement: the same discussion can rescue an initially wrong majority or destroy an initially correct one. Standard final-accuracy evaluations conflate these opposing mechanisms. We introduce an auditable protocol for homogeneous debate on multiple-choice questions (MCQs) that records each run as a transition ledger over collapse, correction, onset, and signed intervention utility. On 6,925 MMLU-Pro debates, the protocol identifies 253 collapses and a parallel correction ledger that changes how interventions should be judged. Replay experiments reveal the central tradeoff: a leave-one-model-out probe-gated freeze prevents 29 collapses but loses 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy. A compact pre-debate 8-probe screen is a triage signal: its unadjusted family-level association with conditional-collapse risk is high (G=7, Spearman rho=0.893, exact two-sided p=0.0123), but initial-majority accuracy is a close comparator (rho=0.821; family partial rho=0.767, p=0.0877), so we do not treat it as calibrated or capability-adjusted prediction. Round-level traces localize many collapses to the first debate round, where early disagreement can precede both harmful cascades and useful recovery. We release replayable schemas, coders, audits, cost cards, and zero-API rebuild scripts so future model-scaffold rows can be compared under the same denominators and signed utility ledger.

Production large language models retrieve and cite web pages alongside generated answers, yet the page-level features that predict citation frequency remain poorly characterised. We present an observational study of approximately 2 million LLM citations from four commercial engines (ChatGPT, Claude, Google AI, Gemini) over six months, joined to 10,000 crawled pages from nineteen B2B SaaS workspaces. Sixty-plus features are tested using a nine-method consensus framework combining mixed-effects regression with domain fixed effects, FDR correction, stability-selection Lasso, double machine learning, generalised additive models, and temporal hold-out replication. Four findings survive all checks. First, prompt-content alignment (Jaccard overlap between page tokens and the full workspace prompt corpus, including non-citing prompts) is the dominant page-level predictor (beta = +0.37, 95% CI [+0.33, +0.41], q ~ 10^-73). Second, the standard AEO checklist (FAQ blocks, structured data, Core Web Vitals) shows positive effects in pooled data that reverse or collapse to zero once domain fixed effects are applied: Simpson's paradox with practical consequences for the AEO literature. Third, domain-level AI authority exceeds the strongest non-alignment page-level feature by a factor of six in mean absolute SHAP value. We release the analytic pipeline as a methodological contribution.

Evaluations of Large Language Models (LLMs) morality typically consider decisions in isolation, thus overlooking whether an individual's unrelated prior conduct influences the model's subsequent choices. This leaves open the question of whether, and to what extent, moral history shapes LLM decisional behaviors. Prior work on human moral decision-making shows that past behavior can influence subsequent moral choices. Building on this observation, we investigate whether analogous effects emerge in LLMs in two complementary ways: at the behavioral level, through the model's observable responses, and at the representation level, through its latent internal representations. We introduce MoralLedger, a framework for studying how an actor's moral history shapes actions for LLMs' behaviors under a fixed decision context. At the behavioral level, we find that prior moral histories systematically alter subsequent choices as a function of their valence and intensity. At the internal representation level, these histories induce a linearly recoverable direction in the residual stream that generalizes to held-out examples. Intervening along this direction on neutral-history prompts produces two-sided intensity-dependent changes in subsequent choices, with effects that are stronger than those induced by prompting alone or by favorable-nonmoral direction. To our knowledge, this is the first demonstration that a latent representation of an actor's prior moral conduct can provide signed inference-time control over a moral decision. Our MoralLedger extends moral evaluation beyond static dilemmas, establishing moral history as both a source of behavioral sensitivity and a causal target for auditing and controlling moral behavior in LLMs.

Mixture-of-Experts (MoE) models repeatedly route tokens to sparse subsets of experts, but conventional routers expose no routing-specific record of how cross-layer influences accumulate. We introduce Scratchpad-Augmented Mixture-of-Experts (SA-MoE), which gives each router access to a low-dimensional persistent state that is not provided to the experts. Learned layerwise writes update this state, and their realized post-update changes exactly decompose the state-mediated contribution to any later routing margin, forming a routing ledger. Across sparsely upcycled SmolLM2- and Gemma-based models and three independent training seeds per architecture, this pathway adds less than 1% analytical forward compute and is strongly used by trained routers: local removal of its router contribution changes the selected Top-2 expert set in 87.6% and 69.9% of decisions, respectively. Relative to a matched latest-write-only control, persistent accumulation increases long-horizon future-routing accessibility by 19.4 and 12.2 percentage points, with positive effects in every seed. More than 90% of absolute ledger contribution comes from non-recent writes in both families, and full-forward suppression of ledger-selected writes changes later routing and output distributions. The ledger is an exact provenance object for the persistent-state pathway, not a complete causal explanation of routing. Sensitivity-aware scores better predict full-forward intervention effects, and post-hoc methods recover related cross-layer attribution without architectural modification. SA-MoE instead makes one routing-specific computational history explicit and directly inspectable within the model's natural forward computation.

Progress in machine learning cannot outpace our ability to verify it. With an explosion in papers today, every scientific claim rests initially on trust in the trainer, leading to uneven evaluation, baselines, and forestalling of reliable progress. Traditionally, the burden of verification falls on the reader, who must reproduce expensive training runs. This strategy is impractical due to an explosion in slop contributions, diversity of methods, and the sheer compute required. We put the burden of proof where it belongs, on the trainer, and in the process also cut the overall cost of verification significantly. We introduce Witnesses, a method for certifying training, data usage and evaluation in a neural network training run. Our key insight is that fast behavioral fingerprints with occasional replay challenges are sufficient for auditing neural network training. Our method is applicable at scale with minimal overhead to the trainer, is cheap for the verifier, rejects bad training runs with amplifiable probability, and allows for exact queries of both data inclusion and exclusion. We test our method on language model training runs from 100M to 2B scales, across DDP and FSDP, and demonstrate this minimal overhead. We also introduce a self-regulating leaderboard of "auto-certified" training runs that enables shared baselines and progress. We invite the community to participate in the leaderboard to improve reproducibility in machine learning.