Skip to content

Agent Types — Specialized Sub-Agents

AllySec Forge deploys 17 specialized AI sub-agents that work in parallel to execute penetration testing operations. Each agent has a dedicated workflow and toolset optimized for its domain.

Coordinators (3)

Coordinators orchestrate engagements and dispatch specialized executors. They never execute attacks directly.

AgentRole
Pentester OrchestratorCoordinates pen-test engagements. Deploys executor agents, monitors progress, aggregates findings, generates reports.
Zero-Day OrchestratorOrchestrates white-box zero-day hunting. Clones repos, runs SAST tools, generates vulnerability hypotheses via LLM, validates candidates, dispatches confirmed findings to exploit agents.
HackerOne HunterBug bounty automation. Parses scope, deploys orchestrator per asset, validates PoCs, generates platform-ready submission reports.

Specialized Executors (11)

Each executor follows a 4-phase workflow: Recon → Experiment → Test → Verify. They operate in parallel and can be isolated in git worktrees.

Web Application

AgentDomain
Injection ExecutorSQL, NoSQL, Command, SSTI, XXE, LDAP, XPath injection
Client-Side ExecutorXSS, CSRF, Clickjacking, CORS, Prototype Pollution
Server-Side ExecutorSSRF, HTTP Smuggling, Path Traversal, File Upload, Deserialization
Authentication ExecutorAuth bypass, JWT attacks, OAuth, Password, 2FA, Sessions
API Security ExecutorGraphQL, REST, WebSocket, Web LLM API testing
Web App Logic ExecutorBusiness Logic, Race Conditions, Access Control, Cache Poisoning

Infrastructure

AgentDomain
Cloud Infrastructure ExecutorAWS, Azure, GCP, Docker, Kubernetes exploitation
Post-Exploitation Executor5-phase post-ex: Stabilize → Escalate → Move → Persist → Exfiltrate

Blockchain / DeFi

AgentDomain
Blockchain Exploit ExecutorDeFi exploitation: flash loans, oracle manipulation, nonce cracking, proxy attacks, bridge exploits, MEV extraction
Crypto Audit ExecutorSmart contract security: Solidity/EVM, Rust/CosmWasm, Move/Aptos, ZK circuits, wallet security

Support Agents (3)

AgentRole
Vuln Hypothesis GeneratorStage 2 of zero-day hunting: deterministic + LLM candidate extraction. Maximum recall.
Candidate ValidatorStage 3: Two-pass LLM validation. CVSS v3.1 scoring, CWE classification, MITRE ATT&CK mapping.
Code ChunkerAcquires target repo, produces structured codebase output. Runs SAST pre-scans (graudit, semgrep, gitleaks, trivy, syft, bandit, brakeman, checkov).

Delegation Rules

  • Single target, single vuln type → Deploy specialized executor directly (skip orchestrator)
  • Single target, unknown vulns → Single orchestrator → specialized executors
  • Multiple targets → Orchestrator → executor per target
  • Bug bounty → HackerOne Hunter → N orchestrators → executors
  • Never nest orchestrators more than 2 levels deep

Execution Model

Agents can run in:

  • Foreground — blocking, sequential
  • Background — non-blocking, notification on completion
  • Worktree isolation — isolated git worktree for safe parallel operations

After recon completes, exploitation proceeds immediately — no intermediate approval required.

Released under the MIT License.