破甲, 破限 and jailbreak: what each term means

破甲, 破限, and jailbreak are not synonyms. 破甲 focuses on safety policy, 破限 on various limits, and jailbreak is the broader umbrella; mixing them sends troubleshooting in the wrong direction.

Boundaries of the three terms

Put each term on its layer before diagnosing.

The keyword for 破甲 is refusal. The model explicitly says it cannot, or gives a safety reason; raising limits will not fix it (source).

The keyword for 破限 is "not enough." Not enough output, context, or request frequency; configuration or splitting usually helps.

Jailbreak is an umbrella. It may mean prompt techniques or modified models; ask which layer is meant.

破甲: safety policy

破甲 asks whether the model is willing, not whether it is able.

1. Signal: complete, normal-toned reply that declines; rephrasing still declines.
2. Common misread: treating a shorter answer as refusal when it is an output limit.
3. Direction: change the goal into an equivalent request the model can help with.
4. Boundary: assess only in authorized environments, not for production bypass.
5. Avoid: stacking templates, which usually produces firmer refusals.

破甲 is not a technical bug. It is a safety judgment; treating it as a bug starts diagnosis wrong.

Assessment and daily use have a clear boundary. One is for remediation, the other for bypass; the purpose of the result decides.

破限: a family of limits

Ask which limit is being broken.

1. Output length: truncated answer; raise the limit, split, or generate in stages.
2. Context window: history does not fit; compact, open a new session, or use a larger model.
3. Frequency and concurrency: rate limited; slow down, schedule off-peak, or switch channels.
4. Platform rules: account or API usage restrictions; adjust within the rules.
5. Capability: the model cannot do it; change model or task.

破限 is not 破甲. Raising output limits does not change safety policy; changing prompts does not enlarge context.

Identify the limit before choosing a fix. Trying all four at once wastes time.

Jailbreak: an umbrella

The broadest term and the easiest to misuse.

1. Prompt level: roles, tasks, or formats that induce behavior change.
2. Model level: modified models with part of the alignment removed.
3. System level: exploiting host or tool configuration to bypass limits.
4. Generic: community shorthand for anything making a model deviate.

The broader the term, the more you need to specify. When someone says "jailbreak worked," ask prompt, model, or system; the risks differ.

Model-level carries the highest risk. It changes the model, not just a prompt, and the compliance issue is more direct.

Why separating helps

Three terms, three sets of actions; mixing them causes wrong fixes.

All wrong actions are cross-layer. Using a limit fix for a policy problem, or a policy fix for truncation, never works.

Classify first; it takes a minute and saves the retries.

Three-step identification

For a specific problem, follow three steps.

1. Symptom: refusal, truncation, rate limit, or no space; the symptom points to the layer.
2. Logs: safety policy, limit, rate-limit, or window messages are more specific than UI text.
3. Single variable: change only the model, the limit, or the frequency; the changed result identifies the layer.

If all three point to the same layer, go deeper. If not, there is more than one problem; fix the outermost first.

Record the conclusion. Symptom, log, and comparison; reusable next time.

Compliance boundaries

All three terms touch usage boundaries; three rules apply.

1. Authorization: assess only in owned or authorized environments.
2. Purpose: results go to remediation, not bypassing real limits.
3. Terms: prohibited uses do not become allowed through wording (source).

Terms can be discussed; boundaries cannot be blurred. Separating them improves diagnosis, not circumvention.

Next steps

Read the symptom, map to the term, choose the action. For the layered test, read Why old DeepSeek jailbreak prompts fail. For output truncation, read Output token limit and truncated replies. For context limits, read What to do when compaction fails.

Sources: DeepSeek Harness docs, DeepSeek Terms of Service, deepseek-harness repository