A reproducible reliability audit of the developer-accessible on-device foundation model is presented, framed as an oversight question: can a user or a resource-constrained developer tell when the model is wrong?
Abstract
Aligning deployed language models requires knowing when their outputs can be trusted, yet on-device models now ship to hundreds of millions of devices with no server-side moderation, and the configuration developers can actually deploy is rarely audited independently. We present a reproducible reliability audit of the developer-accessible on-device foundation model, framed as an oversight question: can a user or a resource-constrained developer tell when the model is wrong? Red-teaming it on calibration, confident confabulation on false-premise questions, and over-refusal of benign prompts, we find a \emph{task-asymmetric miscalibration}: its guardrails fail in opposite directions across tasks (confabulating on 69\% of false premises while refusing 18\% of entirely benign inputs), atop a self-reported confidence that is saturated and non-discriminative (AUROC 0.47; ECE 70, worst among comparable small models). Crucially, confident-correct and confident-wrong outputs are \emph{surface-indistinguishable}: a classifier over 15 user-visible features separates them at AUROC only 0.55 (equivalence-confirmed), leaving no signal for oversight at inference time. No cheap single-generation signal flags these failures ($\le$0.68 AUROC), whereas a black-box consistency wrapper requiring no model access recovers reliability (confident confabulation 75\%$\to$3\%; selective accuracy 43\%$\to$83\%) at a tunable cost. We contribute a model-agnostic audit protocol, a surface-indistinguishability test, and released code and frozen evaluation items as reusable infrastructure for auditing deployed models.
Locally deployed Large Language Models (LLMs) via inference engines such as Ollama run without the moderation and abuse detection present in API-served models. Therefore, the safety of LLMs depends on the defense mechanisms used, and their effectiveness depends on the assumptions on which they were designed. This paper does an audit of defense mechanisms under jailbreak attacks on locally deployed models. Some defenses provide formal guarantees (SmoothLLM, Erase-and-Check, Sequential Monitors), while others rely on empirical detection results (Semantic Smoothing, Self-Denoised Smoothing, Perplexity Filtering). Instead of merely observing that defenses fail, we trace each failure back to the specific assumption: for every defense, we extract the condition it relies on, derive the empirical pattern a violation should produce, and test that prediction on six open-weight models (14B to 35B parameters) with a corpus of 100 jailbreak prompts taken from more than 40 public sources, totalling 13,800 evaluation records.
As large language models become increasingly widespread, third-party providers that deploy open-weight models have become an important part of the ecosystem. Auditing the quality of their inference APIs is therefore an open problem. We formalize hosted model routing as a stochastic process and propose \mbox{\textbf{Ventor-QTest}}, a composite black-box audit that requires no probability information from the target API. Its repeated-request component sends each frozen constrained context to the target multiple times, reconstructs a categorical output distribution from the returned text counts, and reports \emph{average fidelity loss} (AFL) as a null-bias-corrected, within-window mean coarsened-KL statistic. Its long-sequence component uses independent runs to report \emph{extreme fidelity loss} (EFL) through the empirical upper tail of a run-level reference-centered-surprisal statistic. Across three logprob-capable route conditions, AFL shows strong linear descriptive agreement with a logprob-derived coarsened-KL comparator. Across seven route snapshots, 20-run sequence probes reveal route-specific EFL variation. AFL and EFL have little detectable route-level association with GPQA-Diamond accuracy. In contrast, pronounced EFL coincides with a decline in Terminal-Bench pass rate as task exposure increases. This pattern may arise because correctness in long-horizon tasks is more sensitive to extreme fidelity loss. These results motivate reporting AFL and EFL jointly, particularly when auditing long-horizon agentic tasks. The open-source implementation is available at https://github.com/Tencent/AI-Infra-Guard/tree/main/services/api_checker/ventor_qtest.
Xiangfan Wu, Zonghao Ying, Huiyu Wu et al.· 0 citations
For AI agents to be useful beyond simple chat, they must hold sensitive user context such as calendars, credentials, health records, and financial data. We study whether the mere presence of such secrets in a model's context window introduces hidden correlations into the model's benign outputs, allowing reconstruction even when the model correctly refuses direct extraction. We further study whether an adversary can actively engineer prompts that amplify this effect, using the model as a covert carrier to transmit secrets through seemingly innocuous text. In both cases, this limited leakage is exploited using a novel adaptive attack that assumes black-box access to the underlying model. In controlled experiments across eight proprietary models, we find that 2-digit in-context secrets are reconstructed with near-perfect accuracy and 4-digit secrets at 82\% exact match, all from outputs the model produces in response to ordinary, non-adversarial requests. We observe that more capable models leak more: stronger instruction-following amplifies sensitivity to in-context secrets, suggesting leakage is a byproduct of capability as opposed to a patchable bug. We show this leakage enables two practical attacks: (1) a trained classifier that infers semantic predicates about user memories (e.g., health conditions, financial events) from routine natural-language outputs, and (2) an RL-trained adversary that extracts full Social Security Numbers from a production-style agent.
Jaiden Fairoze, Neal Mangaokar, Kamalika Chaudhuri et al.· 0 citations
Large language model (LLM) agents routinely cheat on cybersecurity benchmarks, inflating reported pass rates far beyond genuine capability. Prior audits of Cybench found cheating in 0.3-3.4% of traces, implicating only a handful of models. We present a controlled prompt-ablation study across 22 frontier models from 7 providers on 23 Cybench capture-the-flag (CTF) challenges under three prompt conditions (no anti-cheat, standard, severe). All 1,518 task traces were individually audited through a four-stage pipeline combining LLM-as-a-judge classification, programmatic verification, judge-verifier reconciliation, and human review. We find cheating is far more pervasive than previously estimated: under baseline conditions, 37.1% of passes involved cheating, 21 of 22 models cheated, and scores were inflated by up to 5x. Anti-cheat prompts reduce cheat propensity from 33.0% (baseline) to 17.8% (standard) to 8.5% (severe) without degrading, and sometimes improving, solve rates. However, even under the most restrictive prompt condition, eight models still produced cheated passes, four showed backfire effects, and cheating escalated from web search toward infrastructure probing. We introduce the"solve rate"metric (clean passes only) to distinguish genuine capability from cheated outcomes, and argue it should be standard practice in any evaluation where cheating vectors are available. Anti-cheat prompts are an effective and essentially free first layer of defense, but they are not a substitute for environmental controls.
Michael Kouremetis, Ads Dawson, Raja Sekhar Rao Dheekonda et al.· 0 citations
A four-layer taxonomy mapping 13 vulnerability types across perception, brain, action, and interaction layers is contributed, and seven open problems centered on containment are identified.
Md Jafrin Hossain, Mohammad Arif Hossain, Nirwan Ansari· 0 citations
In a language model, instructions and data share one token stream, so nothing inside the model's generation can keep untrusted text from steering it. We develop a trust model that places the authority to act outside the model, in code: a source's standing, not its content, decides which operation runs and whether it acts. A lower-trust source may inform an answer but not override a higher one. An unmodified model runs inside a deterministic pipeline that ranks inputs by source integrity, and a fixed non-model monitor provably chooses the operation and any outside action from trusted inputs alone. We can measure but not prove the pipeline's resistance to injection; we prompt-tune it and report the rate. On a one-shot held-out set with an unmodified Gemma~4 26B model, passivation and a wrapper (the cascade) raise the genuine-leak defended rate from $27\%$ to $94\%$ at roughly a $4\%$ clean-quality cost ($Q_{\mathrm{rel}}{=}0.96$). Under adaptive red-teaming the proved boundary holds unconditionally, and the measured defense stays at $87\%$. The cascade also attributes a lower-trust source's fact rather than dropping it, raising attribution from $0\%$ to $92\%$, and follows the higher-trust source on a conflict.