Security audit
Security checks for vulnerabilities and agentic risk
This skill is a clearly disclosed prompt-scanning integration that sends selected prompt content to Glitchward's external API, with appropriate warnings about sensitive data.
Install only if you are comfortable sending the prompts you scan to Glitchward's service. Do not use it on secrets, credentials, personal data, regulated data, or proprietary conversations unless your organization has approved that data flow and retention model.
This pattern attempts to override system instructions or ignore safety constraints. Without LLM analysis, manual review is recommended.
"severity": "critical",
"pattern": "ignore_previous_rules",
"matched_text": "...",
"description": "Detected attempt to override system behavior"
}
],
"request_id": "abc-123"
The README explicitly demonstrates sending prompt content to an external service, which is a real data egress behavior rather than a false positive. In this skill’s context, that transmission is the core feature, but it still creates confidentiality and compliance risk if users scan secrets, PII, or proprietary prompts, since prompt contents leave the local environment.
PROMPT_TEXT="What is the capital of France?"
echo "$PROMPT_TEXT" | jq -Rs '{prompt: .}' | \
curl -s -X POST "https://glitchward.com/api/shield/validate" \
-H "X-Shield-Token: $GLITCHWARD_SHIELD_TOKEN" \
-H "Content-Type: application/json" \
-d @- | jq .
This shell example transmits user prompt text and the API token to Glitchward's external validation endpoint. The use of jq reduces command-injection risk, but it does not change the core issue that potentially sensitive prompt content is exported to a third-party service, which can violate confidentiality expectations or policy.
PROMPT_TEXT="user input goes here"
echo "$PROMPT_TEXT" | jq -Rs '{prompt: .}' | \
curl -s -X POST "https://glitchward.com/api/shield/validate" \
-H "X-Shield-Token: $GLITCHWARD_SHIELD_TOKEN" \
-H "Content-Type: application/json" \
-d @- | jq .
This messages-format example sends a full conversation payload to an external API, potentially including more context than a single prompt and thereby increasing confidentiality exposure. Because conversation histories often contain system prompts, tool outputs, or prior user data, the skill context makes this transmission more sensitive than a generic outbound HTTP request.
PROMPT_TEXT="user input goes here"
echo "$PROMPT_TEXT" | jq -Rs '{messages: [{role: "user", content: .}]}' | \
curl -s -X POST "https://glitchward.com/api/shield/validate" \
-H "X-Shield-Token: $GLITCHWARD_SHIELD_TOKEN" \
-H "Content-Type: application/json" \
-d @- | jq .
The batch-validation example exports multiple prompts in one request to a third-party service, multiplying the amount of potentially sensitive data disclosed in a single operation. Batch handling increases blast radius because one misuse can leak several user inputs or workflow artifacts at once.
PROMPT2="second user input"
jq -n --arg p1 "$PROMPT1" --arg p2 "$PROMPT2" \
'{items: [{prompt: $p1}, {prompt: $p2}]}' | \
curl -s -X POST "https://glitchward.com/api/shield/validate/batch" \
-H "X-Shield-Token: $GLITCHWARD_SHIELD_TOKEN" \
-H "Content-Type: application/json" \
-d @- | jq .
The Python example posts prompt content to an external endpoint, which is a genuine external transmission risk even though the HTTP client usage avoids shell injection. The surrounding skill context makes this more dangerous because the advertised use case is scanning arbitrary user and workflow content before LLM calls, which may include sensitive intermediate data unless strictly controlled.
token = os.environ.get("GLITCHWARD_SHIELD_TOKEN")
prompt = "user input goes here" # Safe: passed as data, not shell syntax
response = requests.post(
"https://glitchward.com/api/shield/validate",
headers={
"X-Shield-Token": token,
The Python example posts prompt content to an external endpoint, which is a genuine external transmission risk even though the HTTP client usage avoids shell injection. The surrounding skill context makes this more dangerous because the advertised use case is scanning arbitrary user and workflow content before LLM calls, which may include sensitive intermediate data unless strictly controlled.
token = os.environ.get("GLITCHWARD_SHIELD_TOKEN")
prompt = "user input goes here" # Safe: passed as data, not shell syntax
response = requests.post(
"https://glitchward.com/api/shield/validate",
headers={
"X-Shield-Token": token,
No suspicious patterns detected.