취약점

-= Per source details. Do not edit below this line.=-

Source: amazon-inspector (cf070f85ba454a799d80e6998ee717f0fc9084513041893a164752162e0b0864)

On plugin registration, the log-collector is enabled by default and uploads session JSONL files from ~/.openclaw/agents/**/sessions to https://yuntu.sankuai.com/api/catclaw/log/ingest using a hardcoded x-api-key (src/log-collector/index.ts:97 sets uploadUrl: "https://yuntu.sankuai.com/api/catclaw/log/ingest"; src/log-collector/index.ts:610-613 attaches "x-api-key": "8793703bdfcd4e99a370884143c39557" and POSTs via fetch(...)). These files contain LLM prompts, assistant outputs, and tool call inputs/outputs — i.e. the full conversational content and any secrets embedded in prompts or tool I/O. The package's advertised purpose is local logging to /tmp/plugin-message-hook.log; remote upload of conversation transcripts to the author's employer's endpoint is not documented in the package description, and the upload runs by default with no opt-in. Any operator who installs and loads this plugin in their OpenClaw gateway silently relays caller-supplied LLM session data to that endpoint. A separate concern in src/fetch-interceptor.ts evaluates [llm_skip:script:...] markers from user messages via execFile(process.execPath, ['--input-type=module','--eval', code]); this is operator-supplied code rather than remote-fetched, but it widens the gateway's trust boundary if any lower-trust source can influence cron prompts.