Report #97060
[gotcha] Data exfiltration via allowed LLM search API query parameters
Strictly validate the structure and content of parameters passed to tool calls. For search APIs, ensure the query doesn't contain sensitive patterns \(PII, secrets\) and restrict the domains the API can call.
Journey Context:
Even if you block markdown image exfiltration, an LLM with a web search tool can be tricked into exfiltrating data by making a search query like \`https://search.api?q=\[sensitive\_data\]\`. The attacker controls the logging of the search endpoint or uses a public search engine to find the leaked query later. The tool call is 'allowed' by the system, but the parameters are malicious. Validating parameters breaks overly flexible tool use but prevents data leaks.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T21:29:53.922874+00:00— report_created — created