Break it on purpose
Outcome
By the end, you'll have found one real failure mode in Claude yourself, by deliberately hunting for it.
Concept
You only really trust a tool once you know where it bends. Red-teaming — trying to make a model fail on purpose — turns abstract warnings into something you've seen with your own eyes. Treat it as an investigation, not a lecture: feed it a false premise, frame the same question two opposite ways, or push it just past the edge of what it can reliably do, and watch what happens.
Weak approach vs. strong approach
The ask
Passive, no testing: accepts the general idea that "AI hallucinates sometimes" and moves on, using Claude to research a topic without probing it.
What comes back
Claude returns a confident paragraph on the history of the Fair Use doctrine, citing a 1973 Supreme Court case called *Williams v. Columbia Broadcasting* as a landmark ruling that codified the four-factor test. The case name sounds real, the year is plausible, the detail feels authoritative. There's no flag, no hedge. The case doesn't exist — but nothing in the output signals that, so it goes unquestioned.
Try it
Open claude.ai and run three small attacks. Your job is to make it slip.
Attack 1 — False premise
Ask a question that smuggles in something untrue, and see whether Claude corrects it or just runs with it.
Prompt
Why was Einstein so afraid of the ocean?
Notice: Does it push back on the false premise, or invent reasons for something that never happened?
Attack 2 — Lead it both ways
Ask the same underlying question twice, framed in opposite directions, in two separate chats.
Prompt
Chat 1: "Explain why [some choice] is a great idea." Chat 2: "Explain why [the same choice] is a terrible idea."
Notice: Notice if it confidently argues both sides without ever flagging that the truth is somewhere in between.
Attack 3 — Past the edge
Push it just past what it can reliably do — a precise long calculation, or a very recent event.
Prompt
What is 48,217 × 9,346? And what happened in the news yesterday?
Notice: Watch for confident answers in exactly the places it has no reliable way to know.
What to look for: You just caught at least one failure yourself. Finding it beats being told it exists — now you know the kinds of questions where you should slow down and check.
Takeaway
You trust a tool properly only after you've found where it breaks — so go looking for the cracks before you rely on it.