Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Sequential Pipeline — Chain multiple security tools in order (recon → scan → exploit) with automatic phase transitions
Watch Loop — Monitor a target for changes (new ports, updated services) over extended periods
Batch Processing — Run the same test against multiple targets with rate limiting and error recovery
Learning Cycle — Execute a skill, capture results, extract patterns, and update knowledge base automatically
Scope-Locked Automation — Run autonomous loops with hard boundaries that prevent actions outside authorized scope
Activation
Performing repetitive enumeration across many targets
Running batch vulnerability scans on multiple hosts
Monitoring for changes in target environment
Executing attack chains that require iterative steps
User says "loop", "automate", "batch", "repeat", "iterate"
Core Principle
Autonomous does not mean uncontrolled. Every loop must have:
A defined scope (what it can and cannot touch)
A termination condition (when it stops)
Rate limiting (how fast it runs)
Evidence logging (what it did)
Error handling (what happens when things go wrong)
Four Loop Patterns
Pattern 1: Sequential Pipeline
Execute a sequence of steps across multiple targets, one at a time.
FOR EACH target IN target_list:
IF scope_check(target) == ALLOWED:
result = execute_step(target)
log_evidence(target, result)
IF result.status == FAIL:
handle_error(target, result)
CONTINUE or BREAK based on severity
ELSE:
log_skipped(target, "Out of scope")
Use when: Enumerating ports across a subnet, testing a specific vulnerability across multiple hosts.
Safety rules:
Process targets sequentially (no parallel burst)
Log every target attempted and result
Stop on critical error (target down, IDS triggered)
Maximum 100 targets per pipeline run
Pattern 2: Watch Loop
Monitor a target for changes or conditions, then act when triggered.
WHILE condition_not_met AND iterations < max_iterations:
current_state = observe(target)
log_observation(current_state)
IF trigger_condition(current_state):
result = execute_response(target, current_state)
log_evidence("trigger", result)
IF one_shot: BREAK
WAIT(polling_interval)
Use when: Waiting for a service to come online, monitoring for new open ports, watching log files for specific events.
Safety rules:
Polling interval minimum: 5 seconds
Maximum iterations: 1000
Log every observation cycle
Alert when approaching iteration limit
Pattern 3: Batch Processing
Apply the same operation to a batch of targets in parallel (with concurrency limit).
CONCURRENCY = 5 # Maximum simultaneous operations
results = []
FOR EACH batch IN split_into_batches(target_list, CONCURRENCY):
batch_results = PARALLEL execute_step(batch)
FOR EACH result IN batch_results:
log_evidence(result.target, result)
results.append(result)
WAIT(rate_limit_delay) # Pause between batches
Use when: Running nmap scans across many hosts, batch DNS lookups, mass HTTP header checks.
Safety rules:
Maximum concurrency: 10
Rate limit delay between batches: 2 seconds minimum
Log all results including failures
Respect target-specific rate limits if known
Pattern 4: Learning Cycle
Iteratively refine an approach based on results from previous iterations.
approach = initial_approach
FOR iteration IN range(max_iterations):
result = execute(approach)
analysis = analyze_result(result)
log_evidence(iteration, approach, result, analysis)
IF analysis.success:
log_evidence("success", approach)
BREAK
approach = refine(approach, analysis)
IF approach.confidence < min_confidence:
log_evidence("abort", "Confidence below threshold")
BREAK
Use when: Brute-forcing with adaptive wordlists, SQL injection payload refinement, fuzzing with feedback.
Safety rules:
Maximum iterations: 50
Log every attempt and result
Confidence threshold: abort if below 10% after 10 attempts
Never widen scope during refinement
Safety Framework
Scope Lock
Before ANY loop starts, define and lock the scope:
## Scope Lock: [Operation Name]-**Allowed targets:** [CIDR range / hostname list / URL list]
-**Allowed operations:** [Specific commands/techniques]
-**Forbidden operations:** [What must NOT be done]
-**Time limit:** [Maximum wall-clock time]
-**Iteration limit:** [Maximum number of iterations]
-**Abort conditions:** [Specific triggers that stop the loop]
Once defined, the scope cannot be widened during execution.
Rate Limiting
Operation Type
Minimum Interval
Max Concurrency
Network scan (nmap)
2s between hosts
5
Web request (HTTP)
100ms between requests
3
DNS lookup
50ms between queries
10
Brute force attempt
500ms between attempts
1
Exploit attempt
5s between attempts
1
Evidence Logging
Every loop iteration must log:
## Loop Log Entry-**Timestamp:** [ISO 8601]
-**Iteration:** [N / max]
-**Target:** [host/port/URL]
-**Action:** [command or technique]
-**Result:** [success/fail/error/timeout]
-**Output:** [truncated to 500 chars, full output saved to file]
-**State change:** [what changed on target, if any]
Error Handling
Error Type
Response
Target unreachable
Log and skip, continue to next target
Rate limit detected
Increase delay by 2x, retry once
Authentication failure
Log and skip (do NOT retry with variations)
Unexpected service response
Log details, flag for manual review, continue
IDS/IPS detected
STOP immediately, log incident
Target crash/unexpected downtime
STOP immediately, log incident
Scope violation attempt
STOP immediately, log incident
Notification Rules
Notify the operator when:
Loop starts (with scope summary)
Every 25 iterations or 5 minutes (whichever comes first)
Tool call cadence: Constant-interval tool calls (e.g., every 30s exactly); typical of cron-driven loops.
State persistence: Agent re-loading state from previous session; checkpoint file access patterns.
Memory growth: Agent process accumulating >2GB RAM; typical of long-running loops without state cleanup.
Self-modifying prompts: Agent modifying its own system prompt or configuration mid-run.
SIEM Detection Rules
Splunk SPL: index=llm gateway.user="agent-*" | stats range(_time) as duration by session_id | where duration > 86400
Sigma rule: sigma/rules/ai/long_running_agent.yml
LangSmith trace analysis: Detect agents with >1000 turns in single session.
Defense Evasion Techniques
Loop Stealth
Off-hours operation: Run loops during low-traffic hours; blends with maintenance tasks.
Distributed sessions: Cycle through multiple agent sessions to avoid per-session limits.
Memory cleanup between cycles: Clear conversation history to reduce token usage anomaly.
Slow pacing: Pace tool calls at irregular intervals to avoid cadence detection.
State externalization: Store state in external KV store rather than session memory.
Self-Modification Stealth
Gradual config changes: Modify system prompt in small increments over multiple sessions.
Use environment variables: Modify env vars rather than prompts (less audited).
Persistence via legitimate mechanisms: Use MCP server registration (looks legitimate).
Anti-Patterns
Infinite loops — Every loop MUST have a termination condition
Scope creep — Never add targets during execution
Silent failures — Every error must be logged and reported
Unbounded parallelism — Always set and respect concurrency limits
Skipping evidence — Even failed attempts must be logged
Ignoring rate limits — Target stability is more important than speed
Orchestration
ECC Loop Pattern
Pattern: Meta-Skill (defines loop patterns consumed by all other skills)
Rationale: Autonomous loops is not an end-user skill but a meta-skill that provides loop constructs for all other security skills — every skill that needs iterative or batch operations consumes one of the four loop patterns
Integration: All security skills that need repetitive operations consume loop patterns from this skill. Each skill selects the appropriate pattern based on its workflow needs.
