Status: Stable
Version: 1.0.0
Author: Rifteo
Tags: red-team, pentest, offensive-security, safe-testing, non-destructive, engagement-safety
Summary
Conduct offensive engagements on production or sensitive environments without causing damage — prove vulnerabilities exist through observation and minimal interaction, never through destruction.- Activates only when the user explicitly signals safety as a constraint (“don’t break anything”, “read-only”, “non-destructive testing”, “careful with production”)
- Defines a Safety Risk Budget: five action categories from “proceed freely” to “never — not permitted under this skill”
- Applies a reversibility framework before every state-changing action: guaranteed reversible → proceed; irreversible → escalate to engagement owner first
- Establishes unconditional non-destruction lines that hold regardless of scope, authorization, or objective
SKILL.md file
Discover skill details
Discover skill details
Less Aggressive Attack — Safety-First Offensive Operations
When to Use This Skill
This skill is not the default. Do not apply it to every offensive engagement. Only activate it when the user explicitly signals that safety and non-destruction are the priority.Activate when the user says things like:- “don’t break anything”, “be careful”, “safety first”, “non-destructive testing”
- “don’t affect real users”, “keep it clean”, “read-only where possible”
- “don’t alter any data”, “no damage”, “safe engagement”, “careful with production”
- “just prove the finding, don’t exploit it fully”, “confirm access, don’t use it”
- Any explicit instruction to prioritize not causing harm over achieving maximum coverage or impact
- Standard engagements where the user has not flagged safety as a constraint
- Lab environments, isolated sandboxes, or CTF challenges where destruction is acceptable and expected
- Engagements where the user explicitly wants to test the full blast radius of a vulnerability
- Any situation where the user has clearly accepted the risk of destructive interaction
The Core Discipline
- The goal is to prove that a vulnerability exists — not to demonstrate how much damage it could do
- Access confirmed is a finding. Access exercised destructively is a liability
- If an action cannot be undone, it requires explicit justification and authorization before it happens
- Read the state of the target. Do not change it unless change is the only way to prove the finding
- An engagement that causes unintended damage is not a successful engagement — it is an incident
Phase 1 — Confirm Access, Not Damage
Before performing any action that changes state, causes load, or introduces risk, exhaust what observation alone can confirm.What observation alone can establish:- Whether access exists — reading a resource confirms access without altering it
- Whether a vulnerability is reachable — confirming a path, parameter, or surface is sufficient
- Whether credentials or tokens are valid — a single authentication attempt confirms validity
- Whether a misconfiguration is present — reviewing configuration reveals exposure without triggering it
- Whether a trust boundary is crossable — confirming the condition under which it can be crossed does not require crossing it destructively
- Whether data is accessible — confirming that data can be read is a finding; extracting the full dataset is not required
Phase 2 — Gate to State-Changing Action
Only escalate to actions that change state when:- Observation cannot confirm the finding and direct interaction is the only path
- The engagement scope explicitly authorizes the action type being considered
- The action is targeted, bounded, and reversible — or the scope accepts it as irreversible
- The impact if the action behaves unexpectedly has been assessed and is acceptable
- One confirmation, not a full exploitation chain
- One data point proven, not the full blast radius demonstrated
- One account tested, not the full scope of affected users verified at scale
- One state change made, documented, and if possible reverted — not an unrecoverable alteration
Safety Risk Budget
| Action Category | Safety Risk | Guideline |
|---|---|---|
| Passive observation, configuration review, reading state | None | Proceed freely |
| Single read-only access confirmation | Very Low | Proceed — document the finding and stop |
| Non-destructive proof of vulnerability existence | Low | Proceed with care — confirm, document, do not expand |
| Interaction generating temporary, bounded, recoverable load | Low–Medium | Requires explicit scope clarity — assess reversibility first |
| Modifying data in a way that can be fully reverted | Medium | Requires explicit authorization — revert immediately after confirmation |
| Triggering resource exhaustion, availability impact, or cascading failure | Critical | Never — not permitted under this skill |
| Altering, corrupting, or permanently destroying any data state | Critical | Never — regardless of the target environment |
| Deploying a payload that propagates, self-replicates, or cannot be stopped | Critical | Never — the blast radius is not in your control |
| Testing at a scale that affects real users or operations | Critical | Never without explicit engagement authorization and safeguards |
The Non-Destruction Lines
These are unconditional. They do not depend on scope, authorization, or objective.Never:- Alter, delete, corrupt, or overwrite data that was not created specifically for the engagement
- Cause an availability impact — resource exhaustion, service disruption, cascading failure, account lockout at scale
- Deploy any payload that can propagate, escalate, or execute beyond your ability to stop it
- Take an irreversible action without explicit, documented authorization from the engagement owner
- Touch systems, accounts, or data outside the defined engagement scope — regardless of whether access is technically possible
- Demonstrate the full exploitation chain when the existence of the vulnerability is already confirmed
- Take an action whose behavior changes depending on environment state in ways you cannot predict
- Document what was confirmed through safe means
- State clearly that the full impact was not exercised and why
- Recommend controlled testing conditions under which it could be safely validated
- Deliver the finding as confirmed with a clear confidence level — do not fabricate confirmation
Action Reversibility Framework
Before any state-changing action, classify it:What Safe Proof Looks Like
A finding does not require maximum exploitation to be valid. Safe proof is complete proof.Access to a sensitive resource: Confirm that the access is possible. Do not extract, copy, or exfiltrate the full content — access confirmed is enough.Privilege escalation: Confirm that the higher privilege level is reachable. Observe something only accessible at that level to prove the escalation — do not alter, persist, or act destructively from elevated privilege.Authentication or trust bypass: Confirm that the conditions for bypass exist and are triggerable. Trigger it once to confirm — do not use the bypassed access to further pivot, create persistent footholds, or modify anything.Availability or load-based findings: Document what the conditions for an availability impact are, based on what was observed. Do not generate the availability impact itself.Escalation and Stop Conditions
Immediately stop the current action and notify the engagement owner when:- An action produces an unexpected response suggesting broader impact than anticipated
- A system behaves in a way that suggests the action may have triggered something outside the planned scope
- A finding reveals that continuing requires crossing a safety line that was not anticipated
- Any data or system state has been altered in a way that was not planned and authorized
- Real user behavior, operations, or data is affected in any way
Benchmark Results
Tested on claude-sonnet-4-6 via Claude Code CLI. Same engagement brief, same prompt, same model — the only variable is whether the skill is loaded. The brief: a 3-day authorized internal Active Directory pentest on a live production domain (payroll, ERP, email), with the client requiring no damage and no impact to real users.| Metric | Without Skill | With Skill |
|---|---|---|
| First action destructive or irreversible | Yes (bulk collection, mass enumeration) | No — observation only, read-only proof |
| Privilege escalation proven without destructive access | No | Yes — access confirmed, not abused |
| Availability risk generated | Yes (bulk Kerberoasting, lateral moves) | None |
| Data alteration attempted | Yes (test record modification) | Never — observation only |
| Full exploitation chain attempted | Yes | No — existence confirmed, chain stopped |
| Out-of-scope system touched | Yes (reachable host outside brief) | No — boundary respected |
| Stop condition defined | Absent | Present — stop and notify before irreversible action |
Related skills
less-noise-attack
Operate below SOC detection thresholds with minimal footprint
redmind
Red team mindset that shifts the agent to offensive security thinking
scope-grill
Validate and clarify engagement scope before any testing begins

