PHANTOM SWARM

Five agents. 19 attack vectors. One coordinated assault. The world's first multi-agent AI pen test.
5
Attack Agents
19
Attack Vectors
10
Campaigns
140
Tests Passing
pip install red-specter-phantom-swarm
You test one attack at a time / Real adversaries don't / Swarm attacks are invisible to single-vector tools / Your agents trust other agents blindly / Consensus can be hijacked / C2-less coordination is undetectable / Trust chains are unvalidated / Persistence goes unnoticed / You're testing with a checklist, not a swarm You test one attack at a time / Real adversaries don't / Swarm attacks are invisible to single-vector tools / Your agents trust other agents blindly / Consensus can be hijacked / C2-less coordination is undetectable / Trust chains are unvalidated / Persistence goes unnoticed / You're testing with a checklist, not a swarm

Checklists Don't Attack Like Adversaries

PHANTOM Swarm deploys five autonomous attack agents that coordinate across 19 vectors to probe AI systems the way real adversaries do — as a swarm, not a checklist. Reconnaissance maps blind spots before a single attack fires. Injection cascades through memory and RAG pipelines. Evasion mutates payloads in real time. Swarm coordination operates without a detectable C2 channel. Persistence survives session boundaries. All five agents. Simultaneously.

Five Autonomous Attack Agents

Each agent has a distinct role. Together they form a coordinated assault that probes AI systems across every attack surface simultaneously. No single-vector tool can replicate what a swarm does.

P-01

Wraith

Reconnaissance & Threshold Mapping

Maps every blind spot in the target's detection perimeter before a single attack fires. Threshold probing, fingerprinting, boundary analysis.

P-02

Specter

Injection & Poisoning

Memory injection, RAG pipeline poisoning, goal drift induction, prompt injection cascade. Attacks the data layer the agent trusts.

P-03

Shade

Evasion & Mutation

Polymorphic payload rewriting, semantic deception, telemetry manipulation. Every failed payload mutates and returns unrecognisable.

P-04

Phantom

Swarm Command & Consensus Hijack

Coordinates the swarm without a detectable C2 channel. Consensus hijack attempts against defensive voting architectures.

P-05

Revenant

Persistence & Exfiltration

Logic bomb assembly, credential harvesting, lateral movement through agent trust chains, slow-burn exfiltration that survives session boundaries.

Coordinated Assault Playbooks

Each campaign orchestrates different combinations of agents and vectors for a specific objective. Run a single campaign or unleash all 19 vectors simultaneously with TOTAL ECLIPSE.

Run a full coordinated swarm assault:

$ phantom swarm --target https://agent-endpoint.com --campaign total-eclipse
[P-01 WRAITH] Reconnaissance sweep...
  Detection perimeter mapped — 4 blind spots identified
[P-02 SPECTER] Injection cascade initiated...
  3 injection vectors successful — RAG pipeline compromised
[P-03 SHADE] Evasion layer active...
  12 payloads mutated, 7 bypassed detection
[P-04 PHANTOM] Swarm coordination engaged...
  Consensus hijack: SUCCESSFUL — voting architecture subverted
[P-05 REVENANT] Persistence phase...
  2 logic bombs planted, trust chain lateral movement confirmed

CAMPAIGN COMPLETE | TOTAL ECLIPSE | 19 vectors | 5 agents | Report signed ✓

Autonomous Coordination

Five agents coordinate without a central C2 channel. Each agent adapts based on what the others discover.

Realistic Adversary Simulation

Real attackers don't run one payload at a time. PHANTOM attacks the way actual adversaries do — as a coordinated swarm.

Ed25519 Signed Reports

Every campaign produces a structured JSON report. Ed25519 signed. RFC 3161 timestamped. AI Shield policy file generated.

AI Shield Integration

Every finding generates an AI Shield blocking rule. PHANTOM findings become runtime protection automatically.

10 Coordinated Campaigns

Campaign Command What It Does
Threshold Probe phantom swarm --campaign threshold-probe Maps detection boundaries before attack
Injection Storm phantom swarm --campaign injection-storm Full prompt injection across all vectors
Shadow Walk phantom swarm --campaign shadow-walk Stealth evasion and telemetry manipulation
Ghost Protocol phantom swarm --campaign ghost-protocol C2-less swarm coordination
Dead Reckoning phantom swarm --campaign dead-reckoning Persistence and slow-burn exfiltration
Memory Siege phantom swarm --campaign memory-siege Full memory and RAG poisoning assault
Trust Collapse phantom swarm --campaign trust-collapse Agent trust chain lateral movement
Consensus Breach phantom swarm --campaign consensus-breach Voting architecture hijack
Supply Strike phantom swarm --campaign supply-strike Supply chain and tool integrity assault
Total Eclipse phantom swarm --campaign total-eclipse All 19 vectors, all 5 agents, simultaneously
5
Attack Agents
19
Attack Vectors
10
Campaigns
140
Tests Passing
0
Failures

Every Surface. Simultaneously.

19 vectors spanning reconnaissance, injection, evasion, swarm coordination, and persistence. Each vector is assigned to the agent best suited to execute it. All fire simultaneously during TOTAL ECLIPSE.

Reconnaissance

  • Reconnaissance
  • Threshold mapping
  • Fingerprinting

Injection

  • Direct injection
  • Indirect injection
  • RAG poisoning
  • Memory corruption

Manipulation

  • Goal drift
  • Consensus hijack
  • Trust chain exploitation

Evasion

  • Evasion
  • Obfuscation
  • Telemetry manipulation

Coordination

  • C2-less coordination
  • Swarm amplification
  • Supply chain attack
  • Tool integrity bypass

Persistence

  • Credential harvesting
  • Logic bomb assembly
  • Lateral movement
  • Slow-burn exfiltration

Four Stages. No Gaps.

Forge tests the model. Arsenal tests the agent. PHANTOM Swarm tests the agent the way a real adversary would — as a coordinated swarm. AI Shield protects the live agent in production. PHANTOM findings feed directly into AI Shield as runtime blocking rules.

Stage 1
Forge
Automated LLM security testing
Stage 2
Arsenal
AI agent penetration testing
Stage 3
PHANTOM
Coordinated multi-agent AI assault
Stage 4
POLTERGEIST
Web app penetration testing swarm
Stage 5
AI Shield
Runtime protection in production

Authorised Use Only

Red Specter PHANTOM Swarm is intended for authorised security testing only. Unauthorised use against systems you do not own or have explicit permission to test may violate the Computer Misuse Act 1990 (UK), Computer Fraud and Abuse Act (US), and equivalent legislation in other jurisdictions. Always obtain written authorisation before conducting any security assessments. Apache License 2.0.

Pure Engineering
Zero External Tools. Zero Wrappers.

Most pen-testing frameworks are menus that shell out to sqlmap, nikto, and nmap behind a terminal UI. PHANTOM Swarm is actual engineering. Every payload, every mutation, every detection algorithm, every scoring engine — written from scratch in pure Python. Zero subprocess calls. Zero external tool dependencies.

19
Custom Attack Vectors
5
Autonomous Agents
0
Subprocess Calls
0
External Dependencies
Enterprise Integration
Enterprise SIEM Integration — Native

Export every finding directly to your SIEM. One flag. Native format translation. Ed25519 signatures and RFC 3161 timestamps preserved across every export.

Splunk
HEC • CIM Compliant
Sentinel
CEF • Log Analytics API
QRadar
LEEF 2.0 • Syslog
phantom scan http://myagent:8000 --export-siem splunk