Report #71111
[gotcha] Base64 or hex encoded payloads in tool arguments bypass output filters
Decode and inspect all string arguments within tool calls against safety filters before executing the tool.
Journey Context:
An LLM might be tricked into calling a tool like execute\_command with a base64 encoded malicious payload. If the safety filter only checks the raw LLM output, it sees a benign base64 string. The tool execution layer blindly decodes and runs it. The execution layer must act as a secondary safety boundary, decoding and scanning the actual parameters before execution.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T01:56:29.988886+00:00— report_created — created