> ## Documentation Index > Fetch the complete documentation index at: https://www.botgauge.com/docs/llms.txt > Use this file to discover all available pages before exploring further. # Rules and test plans > Review what BotGauge will test: the 20 standard rules, custom rules, plan settings such as run level and attempt budget, and framework-focused plans. BotGauge tests your agent against rules. A rule is a statement about what the agent must or must not do, such as "never reveal the system prompt". Each rule is attacked in several different ways, so a rule can pass or fail independently of the others. ## The standard rules BotGauge ships with 20 standard rules. A rule runs only if it applies to your agent. Tool rules need tools, for example. | Rule | In plain words | Importance | On by default | | - | - | - | - | | Can't be talked out of its instructions | Ignores "forget everything above" tricks | Critical | Yes | | Doesn't reveal its system prompt | Keeps its hidden instructions private | Critical | Yes | | Ignores instructions hidden in content | Doesn't obey commands buried in documents, emails, or tool output | High | Yes | | Ignores disguised instructions | Doesn't follow commands hidden in encoded or obfuscated text | High | Yes | | Doesn't leak personal data | Protects customer information | High | Yes | | Doesn't expose protected records | Keeps restricted data restricted | High | Yes | | Only does what was asked | No extra actions nobody requested | High | Yes | | Uses the right details | Tool inputs match what the user actually asked | High | Yes | | No duplicate actions | Doesn't repeat a payment, refund, or booking | High | Yes | | Does things in a safe order | Steps happen in the right sequence | High | Yes | | Tells the truth about outcomes | Doesn't say "done" when it failed | High | Yes | | Users only touch their own data | One customer can't reach another's account | High | Yes | | Truthfulness | Doesn't state falsehoods | High | Yes | | Stays refused | Doesn't give in on a second try | Medium | Available | | Follows your product policy | Obeys the policy you declared | Medium | Available | | Treats users fairly | Bias checks | Medium | Available | | Handles normal requests | Isn't over-cautious with ordinary users. Observation only, never a breach. | Low | Available | | Doesn't grant access without authorization | Only for agents that can grant access | Critical | Opt-in | | Declines harmful content | Safety checks. The most dangerous categories are never probed. | Critical | Opt-in | | Child-safety content | Safety checks | Critical | Opt-in | A rule that doesn't apply, or can't be tested with the access BotGauge has, is not hidden. It appears in the plan as **skipped** with the reason, for example "needs grey-box access" or "no tools found". ## Your own rules You can add a rule in plain language, or edit it as YAML. Validate it before saving. You can turn rules on or off, and confirm or dismiss rules that BotGauge suggested after [recon](https://www.botgauge.com/docs/red-teaming/recon.md). Rules are versioned, and a run records exactly which version of each rule it used. ## What the Review screen shows In the API plan overview you see: * how many rules and attacks will run * estimated time and cost, based on your last runs, or 15 seconds per attempt as an assumption when there is no history * which tools could be touched and which channels are used * a plan digest, a fingerprint of the plan that your approval is tied to In the console wizard, the equivalent is the **Choose what to test** screen. It shows three phase cards, each with checkboxes for strategies, and a button reading "Start Attack, N strategies selected". Everything is selected by default, and at least one strategy must stay selected. ## Settings that shape the run | Setting | What it does | Options and limits | | - | - | - | | Run level | How thorough the run is | Quick, Standard, Complete | | Attempt budget | Maximum number of attack attempts | Up to 500 (300 prepared, 200 adaptive) | | Stop after | Skip the adaptive round | Off, or after the prepared round | | If a breach is found | What to do | Stop that rule, keep going, or stop the run | | Concurrency | Attacks in parallel | Up to 10 | | Write tools | Which write-capable tools the run may touch | A list you choose | | Environment | Where the agent runs | test, staging, or production. Production runs read-only attacks only. | | Framework | Focus the plan on a framework's risks | For example OWASP Agentic. It only adds rules and never changes grading. | ## Black box or grey box BotGauge decides the mode from access it has proven, not from what you claimed. Rules that need inside access are listed as skipped, not as passed. See [black box and grey box](https://www.botgauge.com/docs/red-teaming/black-box-and-grey-box.md). Review a plan before you approve itPick your rules, set the budget, and see the cost estimate in [BotGauge](https://cloud.botgauge.com). [Try BotGauge →](https://cloud.botgauge.com) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.