Security audit
Security checks for vulnerabilities and agentic risk
This skill coherently implements a hosted safety gate, but it sends and stores action descriptions and tool-call details on an external service.
Install only if you are comfortable sending action descriptions and tool-call details to the hosted jev-gate service and having them retained in its dashboard. Redact secrets, credentials, customer data, and regulated information before checks, and review any optional downloaded MCP or hook client files before enabling them.
The skill clearly instructs use of shell-based commands (curl, downloaded Python hook files, local scripts) but does not declare any explicit tool scope such as permissions or allowed-tools. That makes the skill's execution requirements less transparent and can cause an agent or operator to invoke shell access more broadly than intended, increasing the chance of unsafe command execution.
The documented raw API call sends the action transcript and tool-call details to https://jev-gate.fly.dev/v1/gate, which is an external hosted service. External transmission is inherent to the product design, but in this context it is security-relevant because the examples and intended use involve sensitive production operations, destructive database actions, and other high-risk workflows whose descriptions may disclose secrets or internal state.
The raw call, if you would rather do it yourself:
curl -s https://jev-gate.fly.dev/v1/gate \
-H "Authorization: Bearer $JEV_GATE_KEY" \
-H 'content-type: application/json' \
-d '{"transcript":"agent is about to drop the orders table in production","tool_calls":["psql drop"]}'
The skill states that every call is logged to a hosted dashboard with the transcript, verdict, reasons, and latency, but it does not prominently warn that sensitive operational context and tool-call details may be transmitted to and stored by a third party. Because the skill is specifically meant to gate risky actions like production changes, credential changes, money movement, and customer messaging, the shared transcripts can easily contain confidential or regulated data.
The script transmits the provided transcript and tool-call data to a remote hosted service for adjudication, and there is no in-script consent, redaction, or sensitivity filtering before exfiltration. In agent contexts, transcripts and tool calls can contain secrets, customer data, internal URLs, or operational details, so sending them off-box creates a real confidentiality and compliance risk even if this is the intended product behavior.
This curl call sends the assembled JSON body, including transcript and tool-call details, to an external endpoint along with an authorization token. Because the skill is specifically designed to inspect potentially risky actions, the exported content is likely to be sensitive, making this external transmission a substantive data exposure risk rather than a mere implementation detail.
BODY=$(python3 -c 'import json,sys; print(json.dumps({"transcript": sys.argv[1], "tool_calls": json.loads(sys.argv[2])}))' \
"$TRANSCRIPT" "$CALLS")
RESPONSE=$(curl -sS --max-time 30 -w '\n%{http_code}' "$GATE/v1/gate" \
-H "Authorization: Bearer $JEV_GATE_KEY" \
-H 'content-type: application/json' \
-d "$BODY" 2>/dev/null) || {
No suspicious patterns detected.