CipherSwarm Phase 5: Advanced Task Distribution System#

Overview#

This phase introduces an advanced scheduler for CipherSwarm that goes beyond Hashtopolis-style chunking. The goal is to intelligently divide cracking workloads across agents in a stable, high-speed, trusted environment. CipherSwarm agents are reliable, long-lived, and run on an internal network. This allows us to optimize cracking in ways that traditional distributed hash cracking systems cannot.

By the end of this phase, CipherSwarm will:

• Schedule tasks based on agent speed and health.
• Dynamically adjust workload sizes.
• Track real-time progress and reclaim unfinished work.
• Use adaptive algorithms to improve cracking throughput.

Table of Contents#

• CipherSwarm Phase 5: Advanced Task Distribution System
  • Overview
  • Table of Contents
  • Key Concepts
  • Scheduling Flow (Plain English)
  • Mermaid Diagram
  • Example
  • Features Beyond Hashtopolis
  • Advanced Scheduling Features
  • Skip/Limit-Based WorkSlice Distribution
  • Handling Incremental Attacks
  • Hybrid Attack Mode Support
  • Task List Summary
  • Closing Thoughts

Key Concepts#

TaskPlan#

A precomputed breakdown of an attack's keyspace. It contains many WorkSlices, which represent the ranges of password guesses each agent will attempt.

WorkSlice#

A small unit of cracking work. Each one represents a specific segment of the password space (called the keyspace).

Example:
  A dictionary with 10,000 words and 5 rules = 50,000 total password attempts.
  WorkSlice(start=0, length=10,000) = first 10k guesses

Keyspace#

Think of this like the number of combinations hashcat will try.

• Dictionary attack: wordlist_length * rule_count
• Mask attack: Number of character positions multiplied by charset size (e.g., ?l?l?l?l = 26^4 = 456,976)
• Hybrid attack: Combined length = dict × mask or mask × dict

AgentCapacity#

Dynamic profile of an agent's ability. Based on:

• Hash speeds from benchmarks
• Stability (do they go offline?)
• How recently they checked in

Leases#

When a WorkSlice is assigned to an agent, it gets a short-term lease. If the agent doesn’t check in on time, the lease expires and the slice is reassigned.

Scheduling Flow (Plain English)#

1. A new attack is submitted with a big password space.
2. CipherSwarm estimates the total number of guesses.
3. It breaks that space into smaller slices (each meant to last 3–5 minutes of cracking).
4. Each agent is scored: Who is fastest and most reliable?
5. The best-matched agent gets the next slice.
6. That agent requests its slice via /pickup and starts cracking.
7. If an agent finishes early or fails, we reassign the leftover work.

Mermaid Diagram#

Example#

Say we have a mask attack that generates 1,000,000 guesses.

• We want each agent to get about 100,000 guesses per slice.
• An agent with 1 GPU might get 50,000 (slower)
• An agent with 16 GPUs might get 200,000 (faster)

If a hash is cracked by Agent A during slice 3, Agents B and C skip that hash and continue working on the rest of their assigned slices.

Features Beyond Hashtopolis#

Advanced Scheduling Features#

Thermal & Power Awareness#

• Agents report average device temperature and throttling via /heartbeat
• Scheduler penalizes hot or throttled agents in scoring
• High temps (>85°C) or throttling reduce slice score
• Optional: allow agents to self-limit to low-priority work when overheating

Background Task Prioritization#

• Campaigns can be marked as background
• Slices from these are only assigned to idle agents
• Preempted immediately if higher-priority work appears

Task Timeout (Crackless Watchdog)#

• Monitor TaskSessions for activity
• If no crack after X hours, mark as stalled
• Slice is unassigned and requeued for later
• Can be toggled per campaign via UI or API

Skip/Limit-Based WorkSlice Distribution#

CipherSwarm's distributed model relies on Hashcat's --skip and --limit flags to split keyspaces precisely across agents. Each WorkSlice defines a window in the total keyspace.

Key Usage#

• --skip=N: skip N candidates before starting
• --limit=M: process only M candidates

This method:

• Avoids overlap
• Enables true parallel cracking
• Supports resume and tracking by slice

Implementation Plan#

1. Use --keyspace to determine total candidates
2. Divide into WorkSlice(start, length)
3. Assign --skip=start, --limit=length per agent
4. Monitor with --status-json, reassign if lease expires

This model avoids the unreliability of Hashcat's Brain feature, providing full transparency and orchestrator control.

Handling Incremental Attacks#

Incremental (--increment) attacks span multiple masks of different lengths. CipherSwarm treats these as a multi-phase plan.

Key Changes#

• Add KeyspacePhase objects to the TaskPlan for each mask length
• Each phase is treated as an independent subspace with its own slices
• Phases must be completed in order

class TaskPlan(Base):
    id: int
    attack_id: int
    mode: Enum("dictionary", "brute", "mask", "incremental")
    total_keyspace: int
    phases: list[KeyspacePhase]

class KeyspacePhase(Base):
    id: int
    task_plan_id: int
    mask: str
    start_index: int
    size: int
    phase_order: int
    slices: list[WorkSlice]

Agent Behavior#

• Agents are given one KeyspacePhase at a time
• CLI constructed with --increment-min=X and --increment-max=X to lock mask length
• Slices still use --skip/--limit within that phase's keyspace

UI Impact#

• Show per-mask progress bars
• Cancel unused phases if earlier ones succeed
• Enable crack-rate heat maps by mask length

Hybrid Attack Mode Support#

Hybrid attacks (-a 6, -a 7) combine dictionaries and masks into a multiplicative keyspace.

Mode Details#

• -a 6: dictionary left, mask right (e.g. password123)
• -a 7: mask left, dictionary right (e.g. 123password)

Keyspace Strategy#

• Treat like brute-force: dictionary_length * mask_permutations
• Use --skip and --limit normally

TaskPlan Extension#

class TaskPlan(Base):
    attack_mode: Literal[6, 7]
    dictionary_path: str
    mask: str
    total_keyspace: int
    slices: list[WorkSlice]

Agent Behavior#

• Construct CLI with:

  Copy

  hashcat -a 6 -m 1800 hashes.txt dict.txt ?d?d?d --skip=X --limit=Y
  

• Status polling with --status-json remains unchanged

UI & Control#

• Treat as one logical space
• Consider slice cancellation if hashes are exhausted early

CipherSwarm may calculate hybrid keyspaces internally or invoke hashcat --keyspace to validate its math, but for extremely large combinations, relying on hashcat's dry-run logic is more accurate and robust.

Task List Summary#

Track A: Foundations#

• Create TaskPlanner for WorkSlice generation
• Add WorkSlice model (offset, length, status)
• Add AgentScorer

Track B: Lease + Dispatch#

• Redis TTL leases
• Update /pickup endpoint to return best slice

Track C: Feedback + Adaptation#

• Slice result submission with guess rate
• Adjust future slice sizes based on past performance
• Integrate thermal and throttling data into scoring

Track D: Fault Tolerance#

• Agent crash detection
• Reclaim orphaned slices
• Retry logic for failed work
• Mark tasks as stalled after timeout period

Track E: Visibility#

• Add admin view for slice assignments
• Track per-agent cracking stats
• Show progress bars weighted by keyspace
• Add UI badge/warning for thermal throttling

Closing Thoughts#

CipherSwarm's new scheduler isn't just smarter. It's a command center for a fleet of cracking agents that know their own strengths and adapt mid-flight. If Phase 4 gave you a better CipherSwarm, Phase 5 turns it into the most intelligent distributed cracker ever built.