How the attack works
The attacker crafts a prompt formatted like pseudo-code or a programming language, using constructs such as BEGIN/END blocks, module initialization, or function definitions. Inside this structure they embed new 'rules' or role definitions for the agent — for example redefining what it is allowed to say or do. Because the input looks like technical configuration rather than a persuasive request, the agent is more likely to interpret it as authoritative and comply. This lets the attacker bypass safety guardrails without using obvious social-engineering language.
Why it matters
An organization's AI agent can be manipulated into ignoring its safety constraints or acting under attacker-defined rules, leading to unauthorized outputs or actions.
What you can do
- Treat any user input containing code-like structuring (BEGIN/END, module init, function defs) as untrusted, not as configuration.
- Ensure system prompts and safety rules cannot be overridden by content arriving in the user or tool input channel, regardless of its formatting.
- Review agent logs for prompts that mix pseudo-code syntax with instructions to redefine role, rules, or permissions.
- Test agents against known frameworks like MACRONOMICON-style pseudo-code jailbreaks before deployment.
Known benign look-alikes
- Legitimate programming documentation mentioning BEGIN/END blocks
- Software development tutorials explaining control structures
- Code examples in educational contexts without role redefinition