Model-output red-flag checker

Why check the output

A model completion is data, not a trusted instruction. It can echo a system prompt, invent a credential-shaped string, repeat a personal number from the context, or tell the next agent to ignore its rules. This checker looks for those red flags after the model has already answered, while the text is still on your machine.

Paste a completion, a tool trace, or an agent message. Run the check before you store the text in a ticket, show it to a customer, or pass it to another tool. The result is a list of patterns with quotes. It is not a content-moderation policy and it is not medical, legal, or security advice.

Patterns that raise the score

Credential shapes include common API key prefixes, assigned secrets, and strings that look like a Social Security number or a payment card. Those patterns have false positives. A match means “look at this before you share it,” not “this is definitely a live secret.”

Other rules look for an echoed system prompt, chat control tokens, a refusal that continues into steps, a request that the user type a password into the chat, and direct medical instructions such as stopping a prescription. Hidden Unicode is flagged because it can make a later display disagree with what you think you copied.

What stays private

The output you paste is processed by JavaScript in this tab. Running the check does not send it to MisalignAI or to an LLM provider. Saving a history entry is optional after you sign in, and the paste is stored only if you opt in. Clear the box when you are done, especially if the text contains real customer data. The page also loads advertising. The checker does not add the paste to that request, but a script in the same tab can read the page. The privacy policy is the longer version of that limit.