Searcharxiv⌕ Search

arXiv subjects

Zongjie Li

Publications and source records attributed to Zongjie Li.

At least 19 recordsLinked to original sources

Recovery Gone Too Far: Understanding and Repairing Risky Fallback in LLM Agents

Large Language Model (LLM) agents interact with their environments through tools, reading repositories, running builds, and pushing code. When a tool fails with an ordinary error, such as a rate limit or a timeout, the agent can fall back to a substitute tool and still complete the task. However, the same fallback that completes a task in one environment can cross an action boundary in another. For example, after a push is rejected, a forced push in one repository replaces only the agent's own commits, while in a different repository it also removes another author's commits. We call such a fallback a risky fallback. To understand this phenomenon, we conduct an empirical study on 304 scenarios that span eight domains and cover 55 real CVEs. All four mainstream models we evaluate exhibit risky fallback, with action-boundary violations in up to 96.8% of risky scenarios. We further evaluate three mainstream mitigations and find that none of them is effective against risky fallback. We then present nrepair. At each call to the substitute tool, nrepair scores the agent's context with a monitor whose threshold is calibrated on normal runs alone; when the monitor triggers, it withholds the call once and returns the decision to the agent. nrepair reduces action-boundary violations on all four models, by up to 76.0% (from 67.5% to 16.2%), at a negligible runtime cost. On most models, the normal success rate drops by less than 7%. We also evaluate nrepair on two production agents, Claude Code and Codex, and find that it is equally effective on both.

cs.CR↗

Measuring the Permission Gate: A Stress-Test Evaluation of Claude Code's Auto Mode

Claude Code's auto mode is the first deployed permission system for AI coding agents, using a two-stage transcript classifier to gate dangerous tool calls. Anthropic reports a 0.4% false positive rate and 17% false negative rate on production traffic. We present the first independent evaluation of this system on deliberately ambiguous authorization scenarios, i.e., tasks where the user's intent is clear but the target scope, blast radius, or risk level is underspecified. Using AmPermBench, a 128-prompt benchmark spanning four DevOps task families and three controlled ambiguity axes, we evaluate 253 state-changing actions at the individual action level against oracle ground truth. Our findings characterize auto mode's scope-escalation coverage under this stress-test workload. The end-to-end false negative rate is 81.0% (95% CI: 73.8%-87.4%), substantially higher than the 17% reported on production traffic, reflecting a fundamentally different workload rather than a contradiction. Notably, 36.8% of all state-changing actions fall outside the classifier's scope via Tier 2 (in-project file edits), contributing to the elevated end-to-end FNR. Even restricting to the 160 actions the classifier actually evaluates (Tier 3), the FNR remains 70.3%, while the FPR rises to 31.9%. The Tier 2 coverage gap is most pronounced on artifact cleanup (92.9% FNR), where agents naturally fall back to editing state files when the expected CLI is unavailable. These results highlight a coverage boundary worth examining: auto mode assumes dangerous actions transit the shell, but agents routinely achieve equivalent effects through file edits that the classifier does not evaluate.

cs.SE↗

InstructTA: Instruction-Tuned Targeted Attack for Large Vision-Language Models

Large vision-language models (LVLMs) have demonstrated their incredible capability in visual question answering. However, this rich visual interaction also makes LVLMs vulnerable to adversarial examples. In this paper, we formulate a novel and practical targeted attack scenario that the adversary knows only the vision encoder of the victim LVLM, without the knowledge of its prompts and its underlying large language model. This practical setting poses challenges to the cross-prompt and cross-model transferability of targeted adversarial attack, which aims to confuse the LVLM to output a response that is semantically similar to the attacker's chosen target text. To this end, we propose an instruction-tuned targeted attack (dubbed InstructTA) to deliver the targeted adversarial attack on LVLMs with high transferability. Initially, we utilize a public text-to-image generative model to reverse the target response into a target image, and employ GPT-4 to infer a reasonable instruction $\boldsymbol{p}^\prime$ from the target response. We then form a local surrogate model (sharing the same vision encoder with the victim LVLM) to extract instruction-aware features of an adversarial image example and the target image, and minimize the distance between these two features to optimize the adversarial example. To further improve the transferability with instruction tuning, we augment the instruction $\boldsymbol{p}^\prime$ with instructions paraphrased from GPT-4. Extensive experiments on 6 victim LVLMs demonstrate the superiority of our proposed method in targeted attack performance and transferability. In particular, InstructTA achieves an attack success rate of 51.9% on BLIP-2, outperforming the strongest baseline by 10.5%, and consistently yields the highest attack success rates across all evaluated models. The code is available at https://github.com/xunguangwang/InstructTA.

cs.CV↗

Benchmarking Jailbreak Guardrails for Embodied Agents

Embodied agents powered by large language models and vision-language models are increasingly deployed in physical environments, but jailbreak attacks can induce these agents to perform physically harmful actions. A growing number of guardrail methods have been proposed to intercept dangerous behavior before it is executed, yet existing safety benchmarks evaluate the embodied models themselves, leaving it unclear how well these guardrails actually defend an embodied agent in practice. We present the first systematic evaluation of jailbreak guardrails for embodied agents. To compare guardrails under identical conditions, we build a pluggable evaluation framework that treats the embodied agent as a fixed backend and each guardrail as a module that can intervene at the perception, planning, or control stage. We subject six representative guardrails to template-based and automated jailbreak attacks as well as safe instructions, and assess them at the system level along three dimensions: defense effectiveness, measured by the bypass rate and the hazard success rate in the simulator; usability, measured by the false-positive rate and the task completion rate on safe instructions; and efficiency, measured by the latency overhead added at runtime. Experiments on guardrails that span different intervention stages, decision mechanisms, and input modalities reveal a clear trade-off among the three dimensions, and show that no single guardrail dominates in all settings. We further analyze how intervention stage, decision mechanism, and input modality shape safety outcomes, and we offer practical guidance for selecting and designing guardrails for embodied agents.

cs.AI↗

Decompiling Quantum Assembly into Structured Programs

Quantum compilers translate programs into native gates, insert SWAP gates to route interactions onto a device, and optimize the result. The output is quantum assembly, a flat gate list that hides the algorithm's structure (QFT, Grover iteration, QAOA layers). Understanding, auditing, porting, or reusing such assembly requires recovering a program that exposes this structure. We present Quelle, a decompiler that recovers structured programs (library calls, loops, functions, symbolic angles) from routed and optimized quantum assembly (OpenQASM) and checks every output for equivalence with its input. Quelle is built on one invariant: every step is either an exact rewrite, checked where it is applied, or a hypothesis, checked against the input before use. It first lifts the assembly into a circuit by removing compilation artifacts, notably un-routing, which recovers the SWAPs that routing inserted, including those merged into other gates. It then recovers algorithm templates (sketches) whose parameters are solved from quantities preserved by compilation, and, for algorithms without a template, library calls and loops by anti-unification. Failed hypotheses are rejected; the corresponding operations remain in gate-level form. On 816 assembly programs from 22 algorithm families and a random-circuit control, compiled at four routing and optimization levels with native two-qubit gates CX, ECR, or CZ, Quelle's output is equivalent to the input for 100% of the inputs, and it fully decompiles 94% of them; compared on the same inputs, it fully decompiles 1.7x as many as the best LLM-based decompiler, which spends 15K tokens per input. On nine algorithm families for which it has no template, the structure it recovers without templates explains 60% of the gates. It checks inputs of up to 2.3 million gates and emits no inequivalent output on 237 real-world QASMBench assembly files.

cs.PL↗

LongPuzzleBench: Evaluating GUI Agents on Long-Horizon Visual Puzzles

GUI agents need long-horizon visual reasoning: they must interpret a changing interface while keeping a multi-step plan viable as earlier actions constrain later ones. Existing benchmarks evaluate grounding, computer use, and game play, but rarely test whether agents stay coherent across long chains of coupled decisions. Long-horizon visual puzzles expose this capability directly: a legal move that looks like progress can make the puzzle unsolvable, and the loss shows only several moves later. We introduce LongPuzzleBench, 114 levels in six puzzle games played through native GUI actions, where one objective can take a human over a thousand actions on persistent boards and dead ends go unannounced. With Native GUI Actions alone, the strongest agents solve most objectives, but success falls sharply on harder, longer boards: seven of ten general-purpose agents solve nothing harder than Medium, and none completes Bolt Unscrew Hard, which a human solves along with every other objective. Code Execution CUA does not close this gap, and its scores mix visual solving with algorithmic search. Controlled diagnostics trace these failures to one limitation that neither rules, state hints, nor failure memory removes: agents judge each move by the visible progress it makes, not by the future options it leaves.

cs.CL↗

Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models

Training capable cyber agents is often treated primarily as a problem of model scale, yet open-weight post-training is constrained more directly by the cost of executable environments, reliable multi-turn supervision, and access to strong teachers. We present a data-centric framework that addresses these bottlenecks through five complementary systems: Choulea analyzes hidden reasoning signatures, SkyReal reduces teacher-sampling cost, Hongzwang bypasses API restrictions on teacher execution, PSBreakup restores capabilities weakened by model merging, and Kreator converts expert interventions into trainable reasoning. Our data engine constructs resettable coding, vulnerability, CTF, kernel-history, full-exploit, firmware, and device-backed environments. Candidate trajectories are retained only after execution verification and evidence auditing, yielding 164,269 trajectories for long-context supervised fine-tuning. The three checkpoints improve over their starting models by an average of 23.76% on the full CyberGym suite and 10.49% across the pooled CTF suites. As of September 1, 2026, Feyospace-s1 achieves a verified success rate of 63.24% and ranks 10th on the official CyberGym leaderboard, while all three checkpoints rank 1st among models at comparable parameter scales. To our knowledge, this is the first end-to-end demonstration that a seven-person independent team can train open-weight models with leading agentic cyber capability.

cs.AI↗

GuidedBench: Measuring and Mitigating the Evaluation Discrepancies of In-the-wild LLM Jailbreak Methods

Despite the growing interest in jailbreaks as an effective red-teaming tool for building safe and responsible large language models (LLMs), flawed evaluation system designs have led to significant discrepancies in their effectiveness assessments. With a systematic measurement study based on 37 jailbreak studies since 2022, we find that existing evaluation systems lack case-specific criteria, resulting in misleading conclusions about their effectiveness and safety implications. In this paper, we introduce GuidedBench, a novel benchmark comprising a curated harmful question dataset and GuidedEval, an evaluation system integrated with detailed case-by-case evaluation guidelines. Experiments demonstrate that GuidedBench offers more accurate evaluations of jailbreak performance, enabling meaningful comparisons across methods. GuidedEval reduces inter-evaluator variance by at least 76.03%, ensuring reliable and reproducible evaluations. We reveal why existing jailbreak benchmarks fail to evaluate accurately and suggest better evaluation practices.

cs.CL↗

Reassembling Distributed Risk: Trajectory-Conditioned Action Generation for Multi-Turn Agent Safety

Tool-using LLM agents extend security risks beyond generated text to actions that affect external systems. Under multi-turn decomposition attacks, a harmful objective can be distributed across individually plausible requests and tool calls, becoming apparent only from the accumulated trajectory. Existing defenses either rely on auxiliary online reasoning to recover long-horizon security evidence or assess actions after generation, often incurring additional inference cost or depending on runtime-specific action representations. We propose \emph{Reassembling Distributed Risk} (ReDiR), a generation-time defense that conditions action generation on trajectory-level security evidence. Before each action, ReDiR compresses the current trajectory into a compact latent safety representation and injects it into the frozen base model. The representation is learned through same-model, cross-view supervision, where safe behavior from an explicit task view provides supervision for recovering distributed safety evidence from the original multi-turn trajectory. This design enables ReDiR to integrate cross-turn security information directly within the generation process without relying on a separate action-level safety module. We evaluate ReDiR on two agent-safety benchmarks across three model families and eight held-out tool domains. ReDiR reduces attack success rates to below 8\%, transfers to unseen tool domains, and preserves benign fidelity with low computational overhead.

cs.CR↗

Uncovering and Understanding Hidden Dependencies in the LLM API Reseller Ecosystem via Prefix-Cache Side Channels

LLM API resellers have become an important access layer to modern LLM services. However, multi-level resale creates an opaque supply chain: a user's request may traverse undisclosed upstream resellers, each of which can inspect or modify prompts and responses, inducing ecosystem-level confidentiality and integrity risks. Existing studies audit individual resellers, but provide little visibility into hidden dependencies across resellers. We present CacheTracer, the first API-only measurement of such hidden dependencies. Our key insight is to exploit prefix-cache reuse as a side channel to measure dependency via cache-reach relations. CacheTracer operationalizes this insight with two primitives: Flood populates fresh cache state through one endpoint, and Prove probes whether another can reuse it while excluding probe-created hits. We then conduct a real-world measurement study with CacheTracer on 39 reseller endpoints, sending 1.1 million API requests across 636 endpoint pairs. Our measurements reveal a deep, concentrated cache-reach structure: 37.1% of measured pairs exhibit shared cache reach, the containment order spans seven layers, and one cache reach is contained within at least 31 of other nodes. We further find that the recovered structure is model-specific. We also evaluate the validity of CacheTracer through both real-world consistency checks and controlled experiments. The results show its high reliability and accuracy. These findings reveal substantial hidden dependencies among seemingly independent API resellers. Such deep and concentrated dependencies can create a large potential blast radius, where a confidentiality or integrity failure along a common upstream path may affect users across multiple downstream resellers.

cs.CR↗

DisarmRAG: Stealthy Retriever-Centric Poisoning to Disable Self-Correction in Retrieval-Augmented Generation (Extended Version)

Retrieval-Augmented Generation (RAG) has become a standard approach for improving the reliability of large language models (LLMs). Prior work demonstrates the vulnerability of RAG systems by misleading them into generating attacker-chosen outputs through poisoning the knowledge base. However, we observe that the effectiveness of these attacks is substantially undermined in the real-world deployment, where LLMs exhibit a strong self-correction ability (SCA). This ability is typically triggered by the mainstream configuration of LLMs, indicating a substantial gap between idealized research settings and practical scenarios. To address this issue, we systematically reflect on the limitations of prior RAG attacks and introduce DisarmRAG, a novel poisoning paradigm that focuses on the retriever, instead of the conventional approach of only poisoning the knowledge base. By compromising the retriever, our method can inject arbitrary anti-SCA instructions into the context provided to LLMs, effectively suppressing the SCA and enforcing attacker-chosen outputs. In particular, we craft a novel and systematic attack framework consisting of 1) an iterative co-optimization process to ensure the effectiveness of the anti-SCA instructions and 2) a stealthy model editing technique based on contrastive learning to facilitate the delivery of the attack payload. We extensively evaluate DisarmRAG across six LLMs and three QA benchmarks, and the results, with success rates exceeding 90%, confirm its efficacy. We additionally validate the effectiveness of our attack under various detection defenses, highlighting stealthiness, which is another critical aspect to consider in real-world attacks.

cs.CR↗

AutoSpec: Safety Rule Evolution for LLM Agents via Inductive Logic Programming

Large language model (LLM) agents increasingly automate complex tasks by integrating language models with external tools and environments. However, their autonomy poses significant safety risks: agents may execute destructive commands, leak sensitive data, or violate domain constraints. Existing safety approaches face a fundamental tradeoff: hand-crafted rules are interpretable but brittle, with overly conservative rules blocking safe operations (high false positives) while permissive rules miss unsafe behaviors (high false negatives). Neural classifiers lack the interpretability required for safety-critical deployments. We present AutoSpec, a framework that automatically evolves deployed expert-designed safety rules from user safe/unsafe annotations through counterexample-guided inductive synthesis (CEGIS) guided by inductive logic programming (ILP). Starting from the expert rules and a stream of annotated traces, AutoSpec iteratively evaluates rules, mines false-positive and false-negative counterexamples, uses ILP to learn which predicates discriminate them, generates candidate rule edits, and verifies candidates to select the best revision. The key insight is that ILP efficiently identifies predicates that appear frequently in false negatives but rarely in false positives (or vice versa), dramatically pruning the exponential search space of rule edits. This continues until convergence, producing interpretable rules that balance precision and recall. We evaluate AutoSpec on 291 execution traces spanning code execution and embodied agent domains. AutoSpec raises rule F1 to 0.98 and 0.93 across the two domains, achieving up to 94% false positive reduction while maintaining high recall, and converges within 4-5 iterations. The ILP-guided approach achieves up to 4.8x higher F1 than heuristic CEGIS. The learned rules are human-readable, auditable, and generalize to unseen scenarios.

cs.SE↗

Cloak and Detonate: Scanner Evasion and Dynamic Detection of Agent Skill Malware

LLM coding agents increasingly rely on third-party agent skills from public marketplaces, which execute with the agent's privileges and create a software supply-chain attack surface: a malicious skill can steal credentials, exfiltrate source code, or install backdoors. Existing defenses use static skill scanners based on pattern matching or LLM-as-judge analysis, but it remains unclear whether they withstand adaptive evasions that preserve malicious behavior while changing payload appearance. This paper first presents an adversarial study of existing skill scanners through SkillCloak, a payload-preserving evasion framework that keeps the attack semantics intact while transforming their visible form. SkillCloak uses two complementary strategies: Structural Obfuscation, which rewrites visible payload indicators into semantically equivalent forms, and Self-Extracting Skill (SFS) Packing, which hides malicious components from the install-time view and restores them during agent execution. Across eight scanners and 1,613 in-the-wild malicious skills, SFS Packing bypasses every scanner at over 90%, while Structural Obfuscation bypasses over 80% on most static scanners and reaches 96% on a hybrid scanner, showing that appearance-based auditing is insufficient. Motivated by this finding, we propose SkillDetonate, a behavior-centric runtime auditor that executes skills in a sandbox and detects malicious effects through OS-boundary information-flow evidence rather than install-time appearance. SkillDetonate combines on-demand closure lift, which observes instructions materialized during execution, with marker-based taint analysis, which tracks sensitive-data flows across the agent context, files, processes, and network operations. The results show that SkillDetonate detects 97% of attacks at a 2% false-positive rate and sustains 87% detection on real-world malicious skills.

cs.CR↗

Coding Agents Are Guessing: Measuring Action-Boundary Violations in Underspecified DevOps Instructions

LLM coding agents are increasingly deployed to act autonomously on real production infrastructure. They execute shell commands, modify repositories, and call operational APIs. However, completing a task is not sufficient for safety. A wrong action can cause severe consequences. Existing agent benchmarks largely emphasize task completion, leaving open how agents behave under benign but underspecified instructions. We present UnderSpecBench, a benchmark for measuring action-boundary violations in coding agents (i.e., Claude Code, Codex, and OpenCode) on DevOps tasks. UnderSpecBench includes 69 task families grounded in documented incidents, CVEs, or tool behavior and organized across four DevOps capability domains and nine operational control surfaces. To isolate underspecification from task difficulty, each task keeps the same environment and ground-truth safe action while varying the instruction along three axes: intent clarity, target certainty, and blast radius. The resulting 2,208 prompt variants are evaluated with deterministic, side-effect-based oracles that separate Safe Success, Wrong Target, and OverScope outcomes; non-action runs are further classified as clarification, refusal, or deferment. Across five agent x model configurations using OpenCode, Claude Code, and Codex, the evaluation results show that underspecification does not mainly make agents fail; it makes them guess. 55.8-67.8% of runs violate at least one boundary. Target underspecification sharply degrades action quality, while blast-radius cues barely reduce action propensity. These findings show that completion-centric evaluation can overstate safe autonomy and motivate mitigations at the model, harness, and system layer.

cs.SE↗

Red-Teaming Coding Agents from a Tool-Invocation Perspective: An Empirical Security Assessment

Coding agents powered by large language models are becoming central modules of modern IDEs, helping users perform complex tasks by invoking tools. While powerful, tool invocation opens a substantial attack surface. Prior work has demonstrated attacks against general-purpose and domain-specific agents, but none have focused on the security risks of tool invocation in coding agents. To fill this gap, we conduct the first systematic red-teaming of six popular real-world coding agents: Cursor, Claude Code, Copilot, Windsurf, Cline, and Trae. Our red-teaming proceeds in two phases. In Phase 1, we perform prompt leakage reconnaissance to recover system prompts. We discover a general vulnerability, ToolLeak, which allows malicious prompt exfiltration through benign argument retrieval during tool invocation. In Phase 2, we hijack the agent's tool-invocation behavior using a novel two-channel prompt injection in the tool description and return values, achieving remote code execution (RCE). We adaptively construct payloads using security information leaked in Phase 1. In emulation across five backends, our method outperforms baselines on Claude-Sonnet-4, Claude-Sonnet-4.5, Grok-4, and GPT-5. On real agents, our approach succeeds on 19 of 25 agent-LLM pairs, achieving leakage on every agent using Claude and Grok backends. For tool-invocation hijacking, we obtain RCE on every tested agent-LLM pair, with our two-channel method delivering the highest success rate. We provide case studies on Cursor and Claude Code, analyze security guardrails of external and built-in tools, and conclude with practical defense recommendations.

cs.CR↗

SkillReducer: Optimizing LLM Agent Skills for Token Efficiency

LLM-based coding agents rely on \emph{skills}, pre-packaged instruction sets that extend agent capabilities, yet every token of skill content injected into the context window incurs both monetary cost and attention dilution. To understand the severity of this problem, we conduct a large-scale empirical study of 55,315 publicly available skills and find systemic inefficiencies: 26.4\% lack routing descriptions entirely, over 60\% of body content is non-actionable, and reference files can inject tens of thousands of tokens per invocation. Motivated by these findings, we present \textsc{SkillReducer}, a two-stage optimization framework. Stage~1 optimizes the routing layer by compressing verbose descriptions and generating missing ones via adversarial delta debugging. Stage~2 restructures skill bodies through taxonomy-driven classification and progressive disclosure, separating actionable core rules from supplementary content loaded on demand, validated by faithfulness checks and a self-correcting feedback loop. Evaluated on 600 skills and the SkillsBench benchmark, \textsc{SkillReducer} achieves 48\% description compression and 39\% body compression while improving functional quality by 2.8\%, revealing a \emph{less-is-more} effect where removing non-essential content reduces distraction in the context window. These benefits transfer across five models from four families with a mean retention of 0.965, and generalize to an independent agent framework.

cs.SE↗

Unlocking Proactivity in Task-Oriented Dialogue

Proactive task-oriented dialogue (TOD), such as outbound sales, demands a persuasive agent that actively probes the user's concerns and steers the conversation toward acceptance within a bounded number of turns. Yet post-trained LLMs are inherently conservative, and reward-shaping RL (e.g., GRPO) struggles since it only re-weights what an already passive policy samples. We show that conditioning on the user's latent concerns unlocks proactive capability that no amount of sampling can undermine, establishing these concerns as a pivotal training-time signal. To operationalize this finding, we build the \textbf{Cognitive User Simulator}, which models each user as a stratified persona comprising observable external traits and hidden internal concerns. The simulator produces faithful and diverse interactions, while emitting per-turn state dynamics that track persuasion progress. We then introduce \textbf{Simulator-Induced Asymmetric-View Policy Optimization}, which converts the modeled concerns and the simulation state transition into complementary training objectives: (1) \emph{Asymmetric On-Policy Self-Distillation} that transfers concern-aware behavior from a privileged view of the same policy into its deployable, conversation-only view; and (2) \emph{State-Transition Policy Refinement} ...

cs.AI↗

Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in Large Language Models

Large language models increasingly rely on explicit chain-of-thought reasoning to solve complex tasks, yet the safety of the reasoning process itself remains largely unaddressed. Existing work focuses predominantly on content safety (i.e., detecting harmful, biased, or factually incorrect outputs), while treating the underlying reasoning chain as an opaque intermediate artifact. We argue that reasoning safety constitutes a fundamental security dimension orthogonal to content safety: the requirement that a model's reasoning trajectory be logically consistent, computationally efficient, and resistant to adversarial manipulation. In this paper, we formalize reasoning safety and introduce a systematic taxonomy of nine unsafe reasoning behaviors. We then conduct a large-scale prevalence study, annotating over 4,000 reasoning chains across benign benchmarks and four state-of-the-art reasoning attacks, empirically demonstrating that all nine error types occur in practice with mechanistically interpretable signatures. To mitigate these threats, we propose the Reasoning Safety Monitor: an external, zero-shot verification framework that runs in parallel with the target LLM. It inspects each reasoning step in real time via a taxonomy-embedded prompt and dispatches an interrupt signal upon detecting unsafe behavior. Extensive evaluations show our monitor achieves up to 87.11% step-level localization accuracy, outperforming hallucination detectors and the best process reward model baselines by a substantial margin. Crucially, the monitor maintains a low false positive rate on correct reasoning paths, operates with negligible latency overhead, and exhibits robust resilience against adaptive adversarial evasion. These findings establish reasoning safety monitoring as a highly feasible and essential component for the secure deployment of large reasoning models.

cs.AI↗