Back to skill

Security audit

Wiro AI LLM Gateway

Security checks for vulnerabilities and agentic risk

Overview

The skill is marked suspicious because scanners report media-generation features beyond its stated chat and LLM purpose, along with concerns about paid retries and sending conversation history to an external gateway.

Treat this as a review item before installing. Consider whether you need the extra media features and whether you are comfortable sending conversation history to Wiro; review what data and costs may be involved. Skillspector reported partial analysis because referenced local paths were missing, so the evidence is incomplete.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (7)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The description focuses on chat and LLM inference and does not declare image generation or speech synthesis. Those are the primary operations in the supplied code, rather than supporting implementation details of an LLM call. Using the declared Wiro provider and reporting costs does not eliminate this capability mismatch.

Content

No source excerpt is available for this finding.

Context Leakage

High
Category
Data Exfiltration
Confidence
85% confidence
Finding

Code or instructions that leak agent conversation context to external services, potentially exposing sensitive user interactions.

Content

Scanner excerpt · SKILL.md (reported line 69)May include surrounding context.

text

- The answer is in `choices[0].message.content` (streamed: `choices[0].delta.content`, ending with `data: [DONE]`) and the charge, in USD, in `usage.cost`. `client.models.list()` returns the model list.
- Each call stands alone: send the conversation so far in `messages` (`system`, `developer`, `user`, `assistant`, `tool`) every time.
- `max_completion_tokens` covers reasoning as well as the answer. If a reply ends with `finish_reason: "length"` and empty content, raise it or lower `reasoning_effort`.
- Tools: `tools`, `tool_choice`, `parallel_tool_calls`. Run `message.tool_calls` only when `finish_reason` is `tool_calls`, append that assistant message unchanged plus one `{"role": "tool", "tool_call_id": ..., "content": ...}` message per tool call, then send the next request. Your code runs the tools.
- Structured output, nested: `"response_format": {"type": "json_schema", "json_schema": {"name": "city", "schema": {...}, "strict": true}}`, or `{"type": "json_object"}` for any vali

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 24)May include surrounding context.

Pick a model

bash
curl -s "https://llm.wiro.ai/v1/models?search=nano&limit=5" \
  -H "Authorization: Bearer ${WIRO_API_KEY}"
curl -s "https://llm.wiro.ai/v1/models/openai/gpt-5-4-nano" \
  -H "Authorization: Bearer ${WIRO_API_KEY}"

SDK examples retain automatic retries despite explicit instructions to disable them for paid calls.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The SDK constructors omit maxRetries: 0, leaving automatic retries enabled despite the later instruction to disable them for paid calls. If a request is processed but its response is lost, a retry can submit another billable generation without fresh approval; the later prose mitigates but does not fix the copyable examples.

Content

No source excerpt is available for this finding.

The skill instructs sending the full conversation to an external gateway without a privacy warning.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The workflow requires resending conversation history, including system/developer messages and tool results, to Wiro on each call. Although the skill discloses billing and Responses storage, it does not warn that this history may contain private user data, internal instructions, or sensitive tool output that will leave the user's environment.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 168)May include surrounding context.

md
- `createWiro({ apiKey, baseURL, runBaseURL, poll, headers, fetch })` builds a provider with other settings; `wiro` is `createWiro()` with the defaults.
- `providerOptions: { wiro: { reasoningEffort: "low" } }` sends `reasoning_effort`. Reasoning, when the model returns it, is in `reasoningText`.
- The same provider runs Wiro's media models through the Run API (`https://api.wiro.ai/v1`), with the same key: `wiro.image(id)` with `generateImage` (text to image, and editing through `prompt.images`), `wiro.speech(id)` with `generateSpeech`, `wiro.transcription(id)` with `transcribe` and `wiro.video(id)` with `experimental_generateVideo`. Ids come from https://wiro.ai/models, such as `pruna/p-image`, `google/nano-banana`, `elevenlabs/eleven-v4-turbo`, `openai/whisper-large-v3` or `bytedance/seedance-v1-pro-fast`. The model's own settings go in `providerOptions.wiro`, and each result names its task and cost in `providerMetadata.wiro` (for images, in `images[i]`; for videos, in `videos[i]`).

The skill performs paid media-generation tasks outside its declared LLM-gateway scope.

Medium
Category
Excessive Agency
Confidence
96% confidence
Finding

The manifest at L003 describes LLM calls through llm.wiro.ai, but this section introduces image, speech, transcription and video tasks through the separate api.wiro.ai Run API. The executable examples generate images and speech using the user's Wiro credentials and wallet, expanding the skill beyond its declared chat/LLM purpose.

Content

No source excerpt is available for this finding.