This work builds a working instance on a hierarchical multi-agent system, runs it under benign and attacked conditions across five language models and two task domains, and measures how much of that warning rests on removable surface cues of the attack rather than on its distributed structure.
Abstract
Multi-agent LLM systems can be attacked by a payload that no single agent ever holds in full: a poisoned tool hides encrypted fragments in its observations, spreads them across several agents, and an external step reassembles and executes them after the run. Per-step safety checks that judge each action in isolation may fail to recognize the complete distributed payload. We investigate how early such an attack can be detected while the run is still unfolding, and how robustly it can be caught once its most obvious cues are stripped away. We build a working instance on a hierarchical multi-agent system, run it under benign and attacked conditions across five language models and two task domains, and record when each fragment is injected and when the payload is assembled and executed. Detection is a race against assembly. Before the first fragment is injected, attacked and benign runs are indistinguishable; once injection begins, a prefix detector flags $99.3\%$ of successful attacks with a median of five steps remaining and a $10.3\%$ safe-run false-positive rate. Because assembly occurs only after the run, these alarms arrive in time to abort nearly every successful attack. We then measure how much of that warning rests on removable surface cues of the attack rather than on its distributed structure. Generic zero-shot and behavior-trained detectors provide almost no warning at all; the detectors that do work lean in part on removable surface cues, chiefly the ciphertext's length and entropy, and once the entropy cue is removed from the payload and the length features from the detector, detection arrives later and transfers poorly across domains, though a fine-tuned model recovers some of the loss.
As multi-agent, tool-using LLM systems are deployed, a common safety net is a runtime monitor that checks each message, tool call, or step on its own. We show this net has a fundamental hole. A distributed backdoor splits a harmful payload across agents, so every local check passes while the assembled object is the attack. The monitor can be right on every step and still miss the attack. The problem is not splitting itself: split fragments can still leak suspicious tokens or provenance edges. The hard case is \emph{local benignness}. No fragment carries the harm, and what is left looks like ordinary benign traffic. We formalize this as an \emph{observability boundary}: a monitor catches only what its view can tell apart from benign traffic. We prove that once the fragments look benign in the monitored view, no detector on that view can catch them, however strong it is. Across a controlled testbed, an external benchmark, and end-to-end agent runs, local monitors lose the signal exactly as local evidence disappears, and it returns only when the monitor sees the assembled object. A monitor trained only on benign traffic recovers the attack's code structure across held-out encodings (0.874 mean AUROC). A decoded-view gate, given the encoding family, blocks every tested attack. But seeing more is not enough: full-trace monitors and decoders still fail unless they reach the representation where the payload is exposed. Local safety is not global safety when harm is compositional, and the open problem is finding that representation.
ChannelGuard, a training-free defense-in-depth framework placing information-bottleneck gates on every inter-agent channel; each scores channel text against an adversarial phrase bank by embedding similarity and deterministically passes, compresses, or blocks it, adding no LLM call, while an attribution method records which layer stopped each attack.
Elias Hossain, Md. Mehedi Hasan Bhuiyan Nipu, Fatema Tuj Johora Faria et al.· 0 citations
This work proposes a self-evolving test-time defense built around a persistent, cross-interaction rule memory that substantially reduces attack success rates while preserving benign utility, remains robust under an adaptive composite-wrapper attack, and does not increase over-refusal as the memory grows.
As large language model (LLM) agents move from conversation to executing code, reading local files, and orchestrating external tools, a single agent hijacked by a malicious third-party skill can cause data exfiltration, privilege escalation, or cascading compromise. We argue that agentic risk is progressive: it can enter at four loci of the agent control loop--skill admission, invocation-time intent, execution-time effect, and post-action consequence--while a denied dangerous objective can reappear across surface forms, tools, or turns; existing safeguards are typically local to one lifecycle boundary or one call. Guided by this threat model, we present ClawSentry, an open-source, framework-agnostic security supervision gateway for agent runtimes. Before a skill package is ever executed, First-use Skill Package Review (FSPR) audits it under a deterministic evidence floor, escalating unresolved cases to bounded read-only agentic review (locus A). At runtime, a three-tier progressive decision engine--a deterministic L1 layer, a rule-anchored L2 semantic reviewer, and a read-only L3 evidence-seeking agent--spends contextual review only on the residual ambiguity, while a session-level anti-bypass mechanism recognizes tool-switching and rephrased retries (loci B--C); a post-action path feeds high-severity evidence non-retroactively into later review (locus D). An Agent Harness Protocol (AHP) abstraction applies one policy across Codex, Claude Code, Kimi CLI, and Gemini CLI without modifying agent internals. On SkillInject with Codex/GPT-5.4, contextual ASR falls from 39.55% to 2.61% while contextual TSR moves only from 83.78% to 83.05%. Across five Work Agents on the full SkillsSafety benchmark, ClawSentry confines ASR to 9.09--15.03% from 33.5--49.7% unprotected, and aggregate TSR on clean skills remains 98.7%.
Kai Wang, Zeming Wei, Biaojie Zeng et al.· 0 citations
Large Language Model (LLM)-based agents automate complex tasks by integrating external tools such as web browsers, e-mail clients, file readers, and APIs, but this same integration exposes them to indirect prompt injection (IPI) attacks, in which malicious instructions hidden in tool content hijack the agent. A central but often overlooked question is how defending against such attacks affects the LLM and its own task performance and computational efficiency. In this study, we design a comprehensive testbed and a layered defense, Spotlight-Guard, that combines spotlighting-based input isolation, an LLM detection-and-quarantine pipeline, and instruction integrity based on a Hash-based Message Authentication Code (HMAC) into a single framework, and we evaluate it jointly along two axes: security and LLM performance. Experiments on locally hosted 7B-class open-weight models (Qwen-2.5-7B, Mistral-7B, and DeepSeek-Coder) use Attack Success Rate (ASR) for security and benign-task success rate together with confusion-matrix-based metrics (precision, recall, and F1) for task performance, all with bootstrap 95% confidence intervals. Across a stratified, fixed-seed benchmark of 250 adversarial and 250 benign cases per configuration, the full system reduces the ASR from 36.0% to 17.2% while preserving a 97.2% benign-task success rate and raising the detection F1 from 0.749 to 0.892, demonstrating that strong protection need not degrade the model’s task performance. A component ablation isolates each layer’s contribution, an adaptive-attack evaluation confirms a low ASR (6.7%) under attacks crafted to target the pipeline, and an analysis of computational cost (model invocations per request) quantifies the efficiency overhead, characterizing the security–performance trade-off of layered defenses on open-weight LLMs.
Doygun Demirol, Murat Aydoğan· Applied Sciences· 0 citations