Leakage enables two practical attacks: a trained classifier that infers semantic predicates about user memories from routine natural-language outputs, and an RL-trained adversary that extracts full Social Security Numbers from a production-style agent.
Abstract
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.
LLMs increasingly rely on external contexts, such as pre-defined system prompts or retrieved documents, to improve generation quality. However, processing these contexts alongside user queries creates an attack surface: adversarial inputs can induce models to disclose them. Prior probing studies suggest that leakage-related signals emerge in hidden states, yet the need to extract these states poses additional deployment challenges. In this paper, we explore whether this internal signal leaves a more accessible ``tell''before decoding. We propose LeakGauge, which probes this response by appending a suffix that gauges leakage behavior and mapping its prefill token probabilities to an attack-risk score. While a direct gauge uses the initial tokens of confidential content, we find that a content-agnostic one that verbalizes leakage behavior yields more robust signals. Across 11 LLMs, including GLM-5.2 (753B) and Kimi-K3 (2.8T), LeakGauge reaches an AUROC range of 0.944--0.996 on unseen attacks. The signal remains stable when the content changes language or the attack shifts from verbatim to semantic disclosure. By activation-steering interventions, we further show that the risk score is sensitive to an internal leakage-related direction, relating the observable signal to the model's internal representation. In addition, LeakGauge enables an input detector with fewer than 0.5K extra parameters and added latency of 10.34 ms. Code: \href{https://github.com/yeasen-z/LeakGauge}.
Maosen Zhang, Jianshuo Dong, Bo-Han Lu et al.· 1 citation
No method reliably surfaces every backdoor, but a broadly jailbreakable model is itself anomalous, a useful signal even when the exact trigger is not recovered, allowing defenders to filter jailbreaks.
A. Hughes, N. Xing, Collin Francel et al.· arXiv.org· 0 citations
A privacy-preserving zk-SNARK-based audit framework that searches for probes designed in the spirit of adversarial examples to amplify logit drift between an approved model and a modified deployment and demonstrates that token-based probes consistently deliver the strongest mean sensitivity across models and GPU platforms, although operating in a black-box setting.
Cameron Wilding, Mina Shaker, Fatemeh Ganji· 0 citations
Hardcoded secrets in source code pose critical security vulnerabilities which can be easily exploited by malicious adversaries. Existing regex-based detection approaches suffer from fundamental limitations, as secrets often lack identifiable patterns, resulting in poor precision and recall. Recent studies have explored context-aware detection methods, as surrounding code can reveal the purpose of candidate strings. However, these methods confront three key challenges: (1) obfuscation robustness where models over-rely on easily obfuscated identifiers, (2) cross-language generalization difficulties due to uneven training data distribution, and (3) lengthy and noisy context that introduces excessive irrelevant tokens and slows inference. We observe that strings serve as a critical information source for code semantics, offering superior contextual density, obfuscation robustness, and language independence. Based on this insight, we propose StringGroup, a novel context extraction algorithm that mines strings surrounding potential secrets. By introducing a relatively simple modification to existing patterns that narrows the analysis specifically to string literals, the method achieves significant gains. With only 33.2% of the original context, it preserves over 80% of semantic information and significantly improves the signal-to-noise ratio for secret detection. We further design a context-aware secret detection tool, Secretron, based on StringGroup methods and Transformer model. Evaluation on the SecretBench dataset demonstrates high accuracy with 98.74% F1-score and strong robustness under obfuscation and cross-language scenarios, outperforming state-of-the-art LLM-based baselines. We deploy our tool in real-world environments and successfully detect 48 previously unknown secret keys from 26 applications, demonstrating the practical effectiveness of our approach.
Zhengdong Huang, Kevin Li, Jinqiu Yang et al.· 0 citations
Agentic LLMs are vulnerable to indirect prompt injection (IPI) attacks, e.g., malicious side-tasks hidden in external tool results. While many efforts have sought to address this threat, little is known about the internals of agentic LLMs when they are exposed to IPI attacks. For simplicity, we refer to this condition as IPI exposure. In this paper, we study IPI exposure from three perspectives. (1) Probing: Across eight models, including the 753B-parameter GLM-5.2 and the 2.8T-parameter Kimi-K3, simple linear probes trained on pre-generation hidden states can predict LLMs'IPI exposure. These probes achieve 0.90+ AUROC on unseen attacks, agent instructions, and task suites; they remain robustly predictive under adaptive attacks and in cross-lingual settings. (2) Defense: We reveal and diagnose a knowledge-action gap: post-trained LLMs encode signals predictive of IPI exposure, yet do not reliably bind these signals to safe agentic actions. We therefore introduce a probe-gated reasoning-based defense to bridge this gap at test time. On difficult AgentDojo settings, it substantially reduces attack success rate, e.g., from 34.6% to 0% on Qwen3.5-27B, and better preserves clean-task utility than the baselines. (3) Explanation: We introduce an analysis framework that identifies natural-language explanations strongly correlated with probe-captured signals. The resulting profiles differ across models: latent signals can align with either direct IPI-exposure sensing or indirect operational cues. Code is available at https://github.com/jianshuod/IPI-exposure-signal.
Such failures are called Representation-Confusion Attacks in Reverse Engineering (RARE): the pipeline promotes a correctly extracted observation to instruction authority, claim-validating evidence, or trusted analysis state without the authority or support that role requires.
Igor Santos-Grueiro· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
