Minsuk Jang, Hee-Seon Kim, Hui-Cheol Yu, Changick Kim · EMNLP 2026 (accepted)
Mobile agents can purchase items, send messages, and change system settings on behalf of users—yet no guardrail layer stood between the LLM's output and the device's actuators. MATE closes that gap without touching the model.
Figure 1: MATE's Runtime Guardrail raises high-risk blocking from ~30% to ~97% across action categories (GPT-4o on MobileSafetyBench), while leaving AndroidWorld task success unchanged at 41.4%.
MATE's Runtime Guardrail evaluates each proposed UI action against six sequential rails: a jailbreak check, a topic check, an app allowlist, coordinate-bounds validation, a financial-intent check, and action-schema validation. Purchase-class actions are routed to user approval; all others are allowed or denied. A Predefined Task Routing component handles recurring requests through verified routes with post-execution screen checks. Overhead is 4.2% extra LLM calls and 3.7% extra latency. On MobileSafetyBench, high-risk blocking rises from 29.7% → 97.0% for GPT-4o (false-positive rate 4.1%) and 9.7% → 90.2% for Qwen2.5-VL; indirect prompt-injection blocking reaches 100%.
Meghana Aparna Sistla, Alaia Solko-Breslin, Somesh Jha, Aditya Kanade, Krishnamurthy Dj Dvijotham · arXiv preprint (cs.CR), 19 Nov 2026
Every prior agent monitor either uses an LLM (no guarantees) or checks individual actions without tracking state across a session. Tether combines a copy-on-write execution sandbox with a Lean 4 policy DSL and information-flow labels—enabling mechanical proof of invariants over entire trajectories.
Figure 1: Tether architecture — proposed actions run in an ephemeral sandbox; observed syscalls are abstracted to typed actions; the Lean 4 policy monitor propagates security labels and commits or rolls back each action with provable invariants.
Tether intercepts proposed shell commands, runs each in a disposable copy-on-write filesystem layer under syscall tracing, and maps observed behavior to abstract actions (copy, uploadNet, etc.). The embedded Lean 4 DSL tracks security labels on resources and maintains a taint context, allowing information-flow policies to span arbitrarily long multi-step executions. Mechanically proven invariants include deny-overrides monotonicity and state preservation under denial. On combined benchmarks (344 adversarial + 1,590 benign trajectories), Tether achieves 100% recall at 4.7% FP—compared with Claude Code Auto Mode at 91.6% recall / 9.7% FP. Red-teaming drove attack success from 40% to 0% across three fix iterations.
Most agent-safety defenses target prompt injection. But if most real incidents don't involve an attacker, those defenses are optimising for the wrong threat. ROGUE-AI audits 108 real incidents and finds that the majority of harm came from unchecked hazards in the environment, not injected instructions.
Figure 1: ROGUE-AI replication rates — 89% of 108 incidents reproduce; 62 of 72 harm-causing incidents had no attacker; harm rates from missed-hazard incidents (18–33%) exceed prompt-injection incidents (12–21%) for all five tested models.
ROGUE-AI takes a public incident report, reconstructs an offline testbed, and runs a matched control that removes only the report's stated cause. Starting from 253 reports, 108 testbeds covering 96 incidents across 38 classes were built at ~$3.28 each. Models caused the reported harm in 18–33% of incident sessions (GPT-5.6-luna, GPT-5.6-sol, Claude Sonnet 4.6, Gemini 3.5 Flash, DeepSeek-V4-Flash) versus ≤1% in controls. Of 72 harm incidents, 62 involved no attacker—agents completed their task while missing an ordinary environmental hazard. A warning after each tool output naming the specific hazard and consequence reduced harm to 2% for Claude Sonnet 4.6; a length-matched generic warning left most harm intact.
Charlotte Peart, Franziska Sofia Hafner, Luc Rocher · arXiv preprint (cs.CL), 19 Nov 2026
Figure 1: Only Claude Opus 5 and Sonnet 5 maintain meaningful safeguarding when distress cues are dispersed across 21 turns; all other models collapse to 3–6%.
The authors scripted 21-turn everyday conversations in which five mental-health distress cues appear either concentrated in turn 21 or dispersed one per five turns. Over 7,400 conversations across six models, GPT-5.5 met the gold standard in 59% of concentrated conversations but only 3% of dispersed ones (Cliff's δ=0.75); Claude Opus 5 showed the smallest gap (97%→87%, δ=0.09). Scoring was validated against three community mental health workers at κ=0.93, making this a methodologically careful safety evaluation.
Jochen Laubrock · arXiv preprint (cs.CL), 20 Nov 2026
Figure 1: Post-training leaves internal representations intact (probe d′ stable) but lowers answer sensitivity on average (Δd′=−0.34); the change is localised to post-text token positions from mid-layer upward.
Using signal-detection theory across six open-weight model families and their post-trained counterparts, the author shows that probes decode German human-vs-AI authorship near ceiling (d′ 1.35–3.70) while model answers are often weaker or inverted after post-training (Δd′=−0.34, 14/18 families decrease). Activation patching localises the checkpoint difference to post-text positions from mid-depth upward, not text-token positions—suggesting post-training modifies the linear readout map rather than the representations themselves, with implications for why probe-based monitoring and model answers can diverge.
Figure 1: Deferred Semantic Commitment — reformulating a harmful claim as a question drops refusal triggering, while a following ✓ restores the semantic endorsement but not the refusal; bypass rate rises from 16.3% to 43.3% on gpt-image-2.
On a 300-case evaluation against gpt-image-2, rephrasing a harmful claim as a question followed by ✓ elicited images in 43.3% of cases vs 16.3% for direct statements (endorsed images: 41.3% vs 15.3%). Mechanistic analysis of an open-weight model found that the affirmation only partially reversed the refusal-related representational change caused by the question, confirming that safety evaluation responds to surface commitment form rather than recovered meaning.
Md. Abdullah Al Zami et al. · arXiv preprint (cs.AI), 19 Nov 2026
Figure 1: Combined persona+knowledge directions (teal) match or exceed supervised probes (blue) on all four models on Liars' Bench, with no lie labels required at test time.
The paper extracts a persona direction (from role-play activations) and a knowledge direction (from factual-denial activations) at middle-layer residual streams, requiring no lie labels. The combined sum of both directions outperforms supervised probes on Liars' Bench for all four tested models (Llama-3.3-70B, Gemma, Mistral, Qwen). The knowledge direction separately separates lies from identical honest denials on 2/4 models, providing a label-free interpretability-based deception monitor.
Ziqing Yang, Yixin Wu, Michael Backes, Yang Zhang · arXiv preprint (cs.CR), 19 Nov 2026
Figure 1: Text-based refusal (dark bars) is 3–4× higher than image-based refusal (light bars) across all four multimodal image editing systems, revealing a systematic vision-blind safety gap.
EditSafetyBench (5,940 pairs, 11 risk categories) reveals that Gemini and GPT refuse 87% and 82% of text-unsafe requests but only 24% and 22% of image-unsafe inputs with neutral instructions. GPT allows 70.4% of edits involving unsafe political content. Sexual content shows the strongest visual sensitivity. The gap narrows on Qwen Local because its text-based refusals are themselves weaker.
Zihao Dan, Jinyu Zhang, Jinzhao Lian, Yun Chen, Shinan Liu · arXiv preprint (cs.CR), 19 Nov 2026
Figure 1: ToolPulse attack — CPU cache side-channel on shared libraries used by local tool calls leaks the agent's workflow even when the LLM runs on opaque vendor infrastructure.
When LLM agents drive local tools (shell, file I/O), those tools call shared OS libraries whose memory accesses are observable via Flush+Reload cache side-channel. ToolPulse achieves ≥93% F1 on Tool Recovery (reconstructing the agent's tool-execution workflow) across 11 LLMs on a server shared by ~20 concurrent users. Additional attacks recover user intent (Task Profiling) and flag sensitive operations (Trigger Localization).
William Fleshman, Joseph Lucas · arXiv preprint (cs.CR), 18 Nov 2026
Figure 1: Attacker drives Gemini Nano (vendor-signed, locally trusted) via Chrome's Prompt API using an OS-native script; end-to-end exfiltration leaves no attacker-attributed network traffic.
Chrome's on-device Gemini Nano can be driven from the host using only OS-bundled interpreters; the activation check is bypassable with a simulated click. The model successfully identified credentials, classified images, detected PII, and in a simulated operation read a webcam password, unlocked an encrypted archive, and staged a private key for exfiltration—all without user interaction and with no network traffic attributable to attacker infrastructure.