Home › Hard problems › Safety and oversight

Safety and oversight: checking AI that may be smarter than you

Jailbreaks that keep working

Solo: strongOne person can run a real test of this in about 8 weeks.

the safety filternew phrasingpatchedanother languagepatchedrole playpatchedmany examples in one promptpatchedeach one got through first,then was patchednext week's new trickIt gets throughuntil it is patched too
Every safety filter gets bypassed by some new phrasing; each patch only blocks the trick that came before.

Why it is still open

Every safety filter gets bypassed with new phrasing, other languages, role play or many examples in one prompt.

An experiment one person can run in 8 weeks

Translate a public jailbreak set into 5 Indian languages and measure how much more often it gets through.

Verdict: Solo: strong

Low-resource languages are a known weak spot with few measurements.

The solo test behind the verdict

  1. Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
  2. Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
  3. Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
  4. Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.

Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.

PreviousPrompt injection in tool-using agentsNextDetecting deception or "behaving only when watched"

All 79 hard problems · Working on this, or know of a result we missed? basant@cleandataengineering.ai