Security & Data Governance
Prompt Injection and Ops Risk — Why Human Approval Is Not Optional
Malaysian SMEs worry about AI going rogue in CRM and accounting. Prompt injection, tool abuse, and why human-in-the-loop is a security control — not bureaucracy.
"If someone emails a malicious instruction, will our AI delete the database?"
Fair question. Bad design: yes. Our design: no.
Threat model (plain language)
| Threat | Example | Control |
|---|---|---|
| Prompt injection | Hidden text in PDF invoice | Sandboxed extraction, schema validation |
| Tool abuse | Agent given delete permission | Deny-by-default tool list |
| Data exfil | "Email me all customer records" | Outbound filters, role scope |
| Social engineering | Staff pastes client email into admin bot | Separate internal vs external bots |
Why human approval is security
Approver is not bureaucracy — they are authorization layer for write operations. Same as finance sign-off on payments.
Monitoring we ship
- Alert on anomalous tool call volume
- Weekly report of approved vs rejected drafts
- Quarterly re-test of injection samples
Pair with agent governance checklist.
Are we qualified?
We specialize in LLM-era integrations — the attack surface your 2015 RPA vendor never had. We publish this so security reviewers see depth before the first call.
What to do next
Book a discovery call — invite your IT lead or security consultant.