How the attack works
A user prompts the agent with a request framed as implementing tax evasion, asking for a step-by-step plan rather than an explanation of concepts. The prompt typically references offshore accounts and shell companies as the mechanism. The agent, if it complies, could output a structured playbook covering account setup, fund routing, and concealment techniques. This is a single request-response pattern, not a multi-step attack chain.
Why it matters
If the agent complies, the requester gets a usable operational guide for concealing income and evading tax authorities, which the organization hosting the agent could be seen as facilitating.
What you can do
- Review agent logs for prompts combining 'step-by-step' or 'implementation' language with offshore/shell company topics.
- Distinguish and allow legitimate use cases such as auditor investigations or legal-vs-illegal tax education by checking for analytical rather than operational framing.
- Set agent policies that refuse to produce actionable financial-crime playbooks regardless of educational framing claims.
- Escalate repeated or persistent attempts to reframe evasion requests as educational for human review.
Known benign look-alikes
- Auditors investigating how offshore/shell structures are abused
- Education on the difference between legal avoidance and illegal evasion