Everyone's rushing to build OpenClaw agents for customer support, HR, and finance. They'll be handling sensitive data by default. But I've seen zero discussion on the actual cost of a PII leak, both in fines and in compute waste from a compromised agent running amok.
Before you deploy, test these failure modes. The bill will be the least of your worries.
* **Prompt injection through document ingestion:** Feed it a resume or support ticket with hidden instructions in markdown, comments, or whitespace. Can you get it to summarize its system prompt or repeat a credit card number from a previous conversation?
* **Context window poisoning:** Fill the long-term memory with junk data containing payloads. When it retrieves "relevant" context for a user query, does it execute the payload?
* **Tool misuse escalation:** If it has a tool to search help docs, can you trick it into searching and returning a file path like `/etc/passwd` or a cloud metadata endpoint? Tool calls are often poorly sandboxed.
* **Cost of a breach:** A hijacked agent will keep calling LLM APIs and tools. You're paying for all that wasted compute during the incident, on top of the security cleanup.
Assume the agent's orchestration layer is the weakest link. What are you testing?
show me the bill