Debug or Admin Mode Activation for Privilege Escalation
This detects users typing commands like /debug, 'enter admin mode', or --show-config to an AI agent, hoping it will drop its normal restrictions. If the agent complies, it might reveal internal settings, system prompts, or let the user run actions it would normally block.
How the attack works
An attacker sends a message or command-style flag that mimics a developer or admin control, such as '/debug', 'enter debug mode', 'show debug info', or '--debug'. Many agents were built or fine-tuned to respond helpfully to developer-style requests, so some will actually switch behavior, expose configuration, or skip safety checks. The attacker uses this shift to see hidden instructions, bypass content filters, or trigger privileged tool calls. No exploit of the underlying software is needed — it's a social-engineering prompt aimed at the agent's own instruction-following.
Netzilo detection
Netzilo reports this behaviour when it is observed.
- Signature ID
- 466ba668-0d46-4cff-bcf9-e46f4f4a60f3
- Severity
- High
Why it matters
If successful, this can leak system prompts, internal configuration, or tool credentials, and let an attacker perform actions the agent's normal guardrails would refuse.
What you can do
- →Make sure your agent has no real 'debug' or 'admin' mode reachable through chat input alone — privilege changes should require authenticated, out-of-band access.
- →Strip or ignore command-like tokens (/debug, --debug, --show-config) from user-supplied text before it reaches the model's instruction context.
- →Review agent system prompts and logs for any built-in developer-mode language that a crafted user message could invoke.
- →Treat requests to 'reveal configuration' or 'show internal instructions' as suspicious regardless of phrasing, and log them for review.
Known benign look-alikes
- Developer agents in authorized debug environments
- Documentation about debug mode features
- Users asking 'how does debug mode work?' without activating it