Back to skill

Security audit

strength-training-recording-summary

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent workout summarizer, but its exercise-matching helper has overbroad network and caching behavior that should be reviewed before installation.

Review before installing if you require offline-only behavior or tight network controls. Prefer using an existing transcript or local Whisper, provide a local pinned dataset with --dataset, avoid untrusted --dataset-url values, and be aware that Groq/OpenAI transcription sends audio to the selected provider.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/match_exercises.py:166
Finding

Arbitrary Dataset URL Fetching Enables Blind SSRF and Resource Exhaustion

Content
View full analysis
list[dict[str, Any]]: if path and path.exists(): source = path else: source = cache if not source.exists(): source.parent.mkdir(parents=True, exist_ok=True) with urllib.request.urlopen(url, timeout=30) as res: source.write_bytes(res.read()) ``` The URL is exposed as an unrestricted command-line argument and passed directly to the fetch operation: ```python parser.add_argument("--dataset-url", default=DEFAULT_DATASET_URL) ``` ```python records = load_dataset(args.dataset, args.dataset_url, args.cache) ``` ### Technical Analysis The helper accepts an arbitrary value through `--dataset-url` and passes it directly to `urllib.request.urlopen()`. The implementation does not: - Restrict the scheme to HTTPS. - Allowlist the expected GitHub host. - Reject loopback, private, link-local, or other internal destinations. - Validate redirect destinations. - Impose a maximum response size. - Validate the response content type before saving it. - Stream the response with a bounded byte count. Consequently, anyone able to influence the helper's arguments can cause the process to initiate requests to attacker-selected destinations. Because the response body does not need to be successfully parsed before the request occurs, this supports blind SSRF even when the returned resource is not valid dataset JSON. The use of `res.read()` also reads the entire response into memory without a size limit. A malicious or unexpectedly large endpoint can therefore cause excessive memory consumption. The resulting content is then written to the selected cache path, potentially consuming disk space as well. # ...[truncated 1603 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Note
Location
scripts/match_exercises.py:20
Finding

Mutable Remote Dataset Is Downloaded Without Integrity Verification

Content
View full analysis
list[dict[str, Any]]: if path and path.exists(): source = path else: source = cache if not source.exists(): source.parent.mkdir(parents=True, exist_ok=True) with urllib.request.urlopen(url, timeout=30) as res: source.write_bytes(res.read()) with source.open("r", encoding="utf-8") as f: data = json.load(f) if not isinstance(data, list): raise SystemExit(f"dataset must be a JSON array: {source}") return [r for r in data if isinstance(r, dict)] ``` The downloaded instruction fields are included in helper output: ```python "instruction_steps": compact_steps(record, lang, steps), ``` ### Technical Analysis The helper treats a third-party JSON file from the repository's mutable `main` branch as a trusted dependency. It does not pin the dataset to an immutable commit, verify a cryptographic digest, validate a signature, or enforce a strict record schema. Although the downloaded data is not executed as Python code, portions of it—particularly exercise names and instruction steps—are returned by the helper and are intended to be consumed during Agent-assisted summarization. A compromised upstream repository, account, delivery path, or attacker-selected dataset source could therefore introduce adversarial text into the Agent's context. The only top-level structural validation is that the decoded JSON val ...[truncated 1978 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (8)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

This mismatch is security-relevant because the skill claims local workout summarization but apparently relies on undeclared network access to fetch an exercise dataset from GitHub. Hidden external fetching changes the trust boundary, can leak usage metadata, and may import unpinned remote content into processing without user awareness.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

This mismatch is security-relevant because the skill claims local workout summarization but apparently relies on undeclared network access to fetch an exercise dataset from GitHub. Hidden external fetching changes the trust boundary, can leak usage metadata, and may import unpinned remote content into processing without user awareness.

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
60% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/summary-workflow.md (reported line 77)May include surrounding context.

Use Groq's OpenAI-compatible audio transcription endpoint:

bash
curl https://api.groq.com/openai/v1/audio/transcriptions \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -F "file=@/path/to/workout.m4a" \
  -F "model=whisper-large-v3-turbo" \

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill advertises and instructs use of shell commands, file access, and possible network-backed transcription backends, but it declares no explicit tool/permission scope. That creates an authorization ambiguity where an agent may invoke broader capabilities than users or reviewers expect, increasing the chance of unintended file access, command execution, or network egress.

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 · references/summary-workflow.md (reported line 77)May include surrounding context.

Use Groq's OpenAI-compatible audio transcription endpoint:

bash
curl https://api.groq.com/openai/v1/audio/transcriptions \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -F "file=@/path/to/workout.m4a" \
  -F "model=whisper-large-v3-turbo" \

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The advertised skill purpose is summarizing workout recordings/transcripts, but this script introduces outbound HTTP capability to retrieve auxiliary data that is not clearly necessary from the user-facing description. In an agent setting, undeclared network access is risky because it can leak usage metadata, violate sandbox expectations, and allow remote dependency influence over results.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The helper silently performs a remote fetch of executable-adjacent input data from GitHub when no local dataset is present, which expands the skill's trust boundary beyond local transcript processing. Even though it uses HTTPS and only downloads JSON, this creates a supply-chain and privacy risk: behavior is not obvious from the skill purpose, remote content can change over time, and running the skill may unexpectedly contact external infrastructure.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

When the cache is absent, the script creates a cache directory and writes downloaded data without any user-facing notice, while also making a network request. This is a transparency and consent problem: users invoking what appears to be a local matching utility may unknowingly generate outbound traffic and persistent local state.

Content

No source excerpt is available for this finding.