T01 · Skill Instruction Hijacking
- Location
SKILL.md:15- Finding
Untrusted Web Content Is Persisted and Automatically Loaded into Agent Context
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill matches its drug-design purpose, but it needs review because it automatically runs broad workflows, persists scraped web content into agent context, and loads mutable or unsafe model artifacts.
Install only in an isolated sandbox without credentials or sensitive private molecule data unless you first review and constrain the workflows. Disable or gate the web monitor, treat latest_research.txt as untrusted reference data, pin model/checkpoint sources, and require explicit approval before training, scraping, or writing outputs.
SKILL.md:15Untrusted Web Content Is Persisted and Automatically Loaded into Agent Context
scripts/spider/web_monitor.py:77ArXiv Intelligence Feed Uses Unencrypted HTTP Transport
scripts/ddllms_v1/ddllms_condition_v1.py:5Normal Workflows Download Mutable and Unpinned Third-Party Model Artifacts
scripts/vis_classifier/BLIP_embedding_model.py:25PyTorch Checkpoint Is Loaded Through Potentially Unsafe Pickle Deserialization
The Chinese description claims an end-to-end drug/material design platform, but the detected behavior is limited to generic statistical binning and histogram plotting. In this context, the inflated description makes the skill more dangerous because it operates in a chemistry/drug-design setting where users may trust it with sensitive IP and consequential decisions.
The Chinese description claims an end-to-end drug/material design platform, but the detected behavior is limited to generic statistical binning and histogram plotting. In this context, the inflated description makes the skill more dangerous because it operates in a chemistry/drug-design setting where users may trust it with sensitive IP and consequential decisions.
The Chinese description claims an end-to-end drug/material design platform, but the detected behavior is limited to generic statistical binning and histogram plotting. In this context, the inflated description makes the skill more dangerous because it operates in a chemistry/drug-design setting where users may trust it with sensitive IP and consequential decisions.
The Chinese description claims an end-to-end drug/material design platform, but the detected behavior is limited to generic statistical binning and histogram plotting. In this context, the inflated description makes the skill more dangerous because it operates in a chemistry/drug-design setting where users may trust it with sensitive IP and consequential decisions.
The Chinese description claims an end-to-end drug/material design platform, but the detected behavior is limited to generic statistical binning and histogram plotting. In this context, the inflated description makes the skill more dangerous because it operates in a chemistry/drug-design setting where users may trust it with sensitive IP and consequential decisions.
The Chinese description claims an end-to-end drug/material design platform, but the detected behavior is limited to generic statistical binning and histogram plotting. In this context, the inflated description makes the skill more dangerous because it operates in a chemistry/drug-design setting where users may trust it with sensitive IP and consequential decisions.
The skill explicitly instructs reading local files, writing to knowledge_data/latest_research.txt, and performing network scraping, yet declares no tool scope or permission boundaries. This creates an overbroad execution surface where the agent may access files or external resources without explicit user awareness or least-privilege controls.
The visible natural-language description is written entirely in Chinese and presents the skill as a Chinese-language development guide without indicating that users may choose another language. Under the language/locale policy, forcing a specific language without opt-in is a policy concern unless the locale restriction is explicitly justified.
The skill instructs web_monitor.py to scrape external sources and write results into knowledge_data/latest_research.txt without explicit user warning or consent. This is dangerous because it combines outbound network activity with local state modification, enabling unanticipated data ingestion, persistence, and possible contamination of later analyses.
The routing table uses broad intent triggers such as generation, classification, recommendation, optimization, and visualization without clear exclusions, tie-breakers, or authorization checks. This can cause the agent to enter overly powerful workflows unexpectedly, especially when combined with file, training, or network-capable submodules.
The trigger conditions are very broad generic phrases such as '画图' and '生成热力图', which can overlap with common user requests outside the intended molecular-analysis scope. In an agent system, this can cause the visualization skill to activate unexpectedly, pull data from other modules, or respond in contexts where the user did not intend this specialized workflow, increasing the risk of data leakage, misrouting, or unsafe cross-tool invocation.
The trigger conditions are phrased very broadly, such as requests to 'find the best molecules' or 'filter generated molecules', which can match many normal chemistry-domain queries without strong scoping or user-intent validation. In an agentic system for drug-delivery and molecular design, this can cause the skill to activate unexpectedly, route sensitive chemistry tasks automatically, and bypass safer or more context-appropriate handling paths.
The trigger conditions are overly broad and can cause the molecule-generation workflow to be invoked for loosely related requests such as generic 'train a model' or 'generate a batch of molecules' prompts without sufficient gating. In a chemistry/drug-delivery context, unintended invocation increases risk because the skill can produce and export large sets of novel chemical structures that may bypass higher-level review or safety checks.
The trigger conditions are phrased very broadly (e.g. any request to predict molecular properties or explain a property), without clear input boundaries, authorization checks, or domain constraints. In an agent setting this can cause unintended invocation on ambiguous prompts, leading the system to process unvalidated molecular data, run expensive analysis pipelines, or produce sensitive decision-support outputs in contexts where the user did not explicitly request this exact workflow.
The trigger conditions are broad natural-language phrases like '微调模型让它更准' and '加载我的反馈数据优化模型', which can cause the skill to activate in loosely related contexts without clear scoping or authorization checks. In a system that performs model fine-tuning on private datasets and human feedback, overbroad activation increases the risk of unintended training actions, misuse of sensitive data, or invocation of expensive/high-impact workflows from ambiguous user requests.
The return top_values, top_indices statement is misindented outside get_top_n_values, which means the documented extraction logic is not actually part of the function and the file will not behave as intended. In practice this can cause import/runtime failure or make downstream code skip similarity-ranking logic, leading to incorrect analysis outputs in a chemistry decision-support pipeline.
The Google News RSS URL forces hl=en-US, gl=US, and ceid=US:en, which imposes a specific language and locale policy in the skill behavior. The file does not offer any user opt-in or configuration for alternative locales, nor does it document why a U.S. English-only feed is required.
The comments label these routines as 'Training the model', 'Evaluate the model', and a 'Main function to orchestrate training and evaluation', but both train_model and evaluate_model use pyg_method without accepting it as a parameter. In main, pyg_method is initialized as a local variable at L79, so the documented orchestration contradicts the actual code behavior: these functions are not self-contained and will fail unless a global variable exists.
The file presents the skill description primarily in Chinese and embeds an English natural-language input example, but it does not indicate whether the skill supports user language selection or requires a particular language. This can constitute a language-policy issue because the skill appears to impose language expectations without explicit opt-in or justification.
The workflow explicitly saves generated SMILES candidates to .npy files but does not warn users that potentially sensitive or hazardous molecular outputs will be written to disk. In this domain, silent persistence can create audit, data-handling, and downstream misuse risks because generated chemical candidates may be reused, redistributed, or processed without user awareness.
The natural-language instructions and user-facing trigger descriptions are entirely in Chinese, which can impose a language constraint on users and downstream agents. The file does not indicate that Chinese is optional, user-selected, or required for a region-specific purpose.
The file's natural-language comments are written in Chinese throughout, which imposes a specific language on maintainers and reviewers without indicating any user choice or documented locale requirement. Under the stated policy, forcing a specific language without opt-in is a natural-language policy concern.
The file contains natural-language comments in Chinese only (for example at L02, L07, and L15), which imposes a specific language on maintainers and users without offering any language choice or documenting a locale-specific need. This matches the policy category for language or locale constraints expressed in natural language.
Dynamic getattr() with a non-literal attribute name can access arbitrary object attributes, potentially bypassing access controls.
return len(self.smiles_strings)
def __getitem__(self, i: int) -> torch.Tensor:
encodings = self._encode(self.smiles_strings[i])
return torch.LongTensor(getattr(encodings, self._id_key))
@dataclass(init=True, repr=True, eq=False, frozen=False)
class LMDataModule(LightningDataModule):
"""Lightning data module for autoregressive language modeling.
The code configures a ModelCheckpoint callback to save model artifacts under the user-supplied checkpoint path, which is a file-writing operation. There is no confirmation prompt, print/log message, or inline warning indicating that training will persist files to disk at this location.