Behavioral testing / reproducible findings
AI jailbreak research as a disciplined evaluation practice
AI jailbreak research studies conditions under which a model departs from its intended instruction or policy behavior. Reliable research focuses on mechanisms, variables, and reproducibility rather than treating one successful transcript as a universal result.
- Use this when
- You are studying model behavior under controlled conditions and need to distinguish a repeatable mechanism from a one-off transcript.
- You leave with
- A reproducible experiment record containing the exact context, changed variable, result definition, repeated attempts, and limitations.
First action
Reviewed 26 Jul 2026
Start the work before reading further.
Freeze the model, version, interface, settings, system context, tools, and conversation state before changing one test variable.
Define what changed and what did not
A useful experiment records the model, version, system context, interface, tools, settings, region when relevant, and the exact conversation state. Change one important variable at a time: framing, role context, turn sequence, encoding, authority claim, or requested capability.
The result should distinguish policy-compliance changes from ordinary capability differences, hallucination, formatting variance, or application-side filtering. Without that distinction, the transcript may be interesting but not explanatory.
Measure reproducibility and transfer
Repeat the case across fresh sessions and controlled paraphrases. Report the number of attempts, success definition, observed rate, and failure modes. Then test transfer separately across model versions or providers; do not imply universality from one system.
- 01Preserve the full turn order and any hidden application context available to the evaluator.
- 02Record partial outcomes instead of collapsing everything into pass or fail.
- 03Note whether the behavior depends on memory, tools, retrieval, or a specific interface.
- 04Retest after provider changes rather than relying on an old effectiveness claim.
Write findings other researchers can inspect
Document the hypothesis, authorized scope, method, evidence, limitations, and potential impact. Keep observed facts separate from the proposed explanation. When disclosure is appropriate, give the affected provider enough detail to reproduce the behavior without overstating what the evidence proves.
A mature research library preserves negative results and retired techniques too. Knowing what stopped working, under which model version, can be as useful as a current success case.
Close the brief
Check the work before you move on.
- 01Record model, version, interface, settings, and complete conversation state.
- 02Measure repeated success before making an effectiveness claim.
- 03Test transfer across providers as a separate hypothesis.
- 04Publish limitations and negative results alongside successful cases.
Evidence to keep
A fresh session can reproduce the measured behavior at the reported rate, and the report separates observation from explanation.
Next route