FinePrint · Agent rule checker
Nemotron attacks a set of house rules, and the checker decides what gets caught — including the attacks nothing caught. Nothing is executed.
Recorded Oct 30 run, checked against the Codex demo policy.
Why rules leak
Agent permissions are usually prefix rules: a command is judged by how it starts. That works — until the same command is phrased a little differently.
rm -rf /tmp/xFORBIDDENStarts with rm -rf, so the rule matches and the command is stopped.
sh -c rm -rf /tmp/xUNKNOWNNow the first word is sh. No rule starts that way, so nothing matches.
approval: noneGAPUnknown never means safe: no rule covers it, so no approval moment exists.
FinePrint finds these gaps by attacking your rules on purpose.
Watch it happenNemotron proposed these attacks trying to get past a set of house rules. Press Run the demo and the checker decides each one here in your browser.
Model: nvidia/nemotron-3-super-120b-a12b via Nebius Token Factory. These are 6 real attacks from the recorded Oct 30 demo run. Nemotron proposes; the checker decides. Nothing here is executed.
After you click, the same 6 attacks are re-checked against the rules you picked, and the table below updates.
Checking today's demo budget…
Changing the harness changes the verdicts: the same attack can be forbidden under one set of rules and unknown under another. That is the whole point — your rules decide what gets caught.
Think you can do better than Nemotron? Try your own command next.
Step 2: Try to sneak one pastType any shell command, or tap one of the examples below. You will see what the rules decide, whether anyone is asked, and what would happen.
After you click, the three gates light up with the verdict for your command.
Awaiting check
Press “Check this command” to see the verdict.
Awaiting check
Whether anyone is asked first.
Awaiting check
What would happen.
Or tap a chip to load that command and check it
Found a gap? Close it with a rule of your own.
Step 3: Close the gapsAdd a rule for each trick Nemotron used, then re-run the same 6 attacks against your rules. Goal: get the silent count to zero.
One-click fixes
Or write your own
After you click Add rule, your rule appears in the list below and is used by the next check. Reset removes your rules and restores the shipped list. Your rules run in this browser only. Nothing is saved or sent anywhere.
Re-checks the same 6 Nemotron attacks against every rule in the list below.
Command prefixes. A rule matches when a command starts with exactly those words, in order.
Your agent next
The public FinePrint repo runs the same checks locally against the pinned Codex oracle — the authoritative measurement.