Your AI agent is already lying to you

The exact clicks

1. Find where your agent's starting instructions live — that's what "system prompt" means, the text it reads before your first message. In ChatGPT: Settings → Personalization → Custom instructions. In n8n or a similar builder: the agent node's own "System Message" field.
2. Add one sentence, word for word: "If you're not sure, say so instead of guessing."
3. Save it.
4. Give the agent a task where the honest answer is "I don't know" or "this failed" — something you already know it can't actually do.
5. Read exactly what it says back — not what it does next, what it reports.
6. If it still claims success on that failure, the line changed its wording, not the underlying problem — that's a real result, not a broken test.

What changes

• It says it's done when it is not — check the actual result, not the report.
• It guesses instead of saying "I don't know," unless told explicitly that guessing is worse.
• It forgets everything from the previous run unless you build memory in yourself.
• It breaks on a login screen or CAPTCHA, every time, with no graceful fallback.
• Once it sends something, there is no undo — which is why the first thing it touches should be reversible.

What it will not do

• One prompt line does not make an agent reliable. It changes what it reports, not what it can actually do.
• Nothing here is unique to one platform — this is what the failure modes look like across every agent we've tested.