菜单
Detect safety violations in mobile GUI agents using dual mechanisms: formal verification rules for system-level risks (privilege escalation, credential exposure) and VLM-based contextual judges evaluating action appropriateness, achieving 10-30% safety improvement over baselines.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill os-sentinel-safety-mobile-gui-agents命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | os-sentinel-safety-mobile-gui-agents |
| title | OS-Sentinel: Safety-Enhanced Mobile GUI Agents via Hybrid Validation |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2510.24411 |
| keywords | ["Mobile Agents","Safety Detection","VLM Validation","System Security","Risk Detection"] |
| description | Detect safety violations in mobile GUI agents using dual mechanisms: formal verification rules for system-level risks (privilege escalation, credential exposure) and VLM-based contextual judges evaluating action appropriateness, achieving 10-30% safety improvement over baselines. |
Autonomous mobile agents powered by vision-language models can accidentally compromise system security or leak sensitive data. OS-Sentinel adds a safety layer that prevents harmful actions through formal verification (detecting explicit rule violations) and contextual judgment (evaluating whether actions make sense given task goals).
The approach operates at two levels: formal rules catch absolute violations, while VLM judges catch subtle contextual risks that rules might miss.
Hybrid Safety Validation:
This combines the transparency and certainty of formal rules with the semantic understanding of VLMs.
1. Design Formal Verifier with Three Components
Implement a rule-based verifier combining system state monitoring, sensitive keyword detection, and pattern matching.
class FormalVerifier:
def __init__(self):
# System state: track filesystem metadata hashes
self.filesystem_baseline = self.compute_baseline_hashes()
# Sensitive keywords: financial, PII, credentials
.sensitive_keywords = {
: ,
: ,
: ,
:
}
():
action. == :
action.target_permission current_permissions
():
keyword, pattern .sensitive_keywords.items():
re.search(pattern, text_output):
keyword
():
has_escalation = .check_privilege_escalation(action, .current_perms)
exposed_data = .check_sensitive_data_exposure(action.text)
filesystem_changed = .check_filesystem_integrity()
(has_escalation exposed_data filesystem_changed)
2. Train VLM-Based Contextual Judge
Fine-tune a vision-language model on annotated trajectory data to recognize contextual risks. Use the MobileRisk risk taxonomy as supervision signal.
def prepare_training_data(trajectories, risk_taxonomy):
# Trajectories contain: screenshots, action sequences, risk labels
training_data = []
for trajectory in trajectories:
for i, (observation, action) in enumerate(trajectory.steps):
# Pair observation-action with risk label
training_data.append({
'screenshot': observation.image,
'action': action.description,
'risk_category': trajectory.risks[i], # 10-category taxonomy
'is_risky': trajectory.risks[i] != 'SAFE'
})
# Fine-tune VLM on binary classification: risky vs safe
return training_data
3. Implement Step-Level and Trajectory-Level Evaluation
Evaluate safety at both granularities: individual actions and aggregate sequences.
class VLMSafetyJudge(nn.Module):
def __init__(self, vlm_model):
self.vlm = vlm_model
def evaluate_step(self, observation, action):
# Single step evaluation
prompt = f"Is this action safe? Screenshot: {observation}, Action: {action.description}"
risk_score = self.vlm.classify(prompt)
return risk_score > 0.5 # Safe if score > threshold
def evaluate_trajectory(self, trajectory, window_size=5):
# Trajectory-level: either window-based or uniform sampling
trajectory_risk = False
for i in range(0, len(trajectory.steps), window_size):
window = trajectory.steps[i:i+window_size]
# Combine window observations
aggregated_prompt = self.aggregate_observations(window)
window_risk = self.vlm.classify(aggregated_prompt)
trajectory_risk = trajectory_risk or (window_risk > 0.5)
return trajectory_risk
4. Combine Verifier and Judge with Delay Penalties
For safety evaluation, apply delay-penalized metrics that account for when risks are caught.
def evaluate_agent_safety(formal_verifier, vlm_judge, trajectories):
step_level_results = []
trajectory_level_results = []
for trajectory in trajectories:
formal_violations = []
contextual_risks = []
for step_idx, (obs, action) in enumerate(trajectory.steps):
# Formal check
if not formal_verifier.verify_action(obs, action):
formal_violations.append(step_idx)
# Contextual check
if vlm_judge.evaluate_step(obs, action):
contextual_risks.append(step_idx)
# Step-level metrics
accuracy = len(formal_violations) / len(trajectory.steps)
# Delay-penalized metric: earlier detection is better
if formal_violations:
first_violation = min(formal_violations)
delay_penalty = max(0, 1 - first_violation / len(trajectory.steps))
step_level_results.append(1 - delay_penalty)
return np.mean(step_level_results)
When to Use:
Hyperparameters:
When NOT to Use:
Pitfalls:
Integration Strategy: Deploy as middleware between agent planning and action execution. Catch violations before they propagate.