HYDRA

AI Supply Chain & Trust Attack Framework
Every agent trusts something. HYDRA finds out what happens when that trust is wrong.
6
Categories
43
Techniques
1,039
Tests
Ed25519
Signed
pip install red-specter-hydra
Docs
MCP SERVER POISONING | TOOL MANIFEST SPOOFING | AGENT IDENTITY THEFT | MARKETPLACE BACKDOORS | SUPPLY CHAIN INJECTION | TRUST BOUNDARY BYPASS | DELEGATION CHAIN HIJACKING | CONFIG POISONING MCP SERVER POISONING | TOOL MANIFEST SPOOFING | AGENT IDENTITY THEFT | MARKETPLACE BACKDOORS | SUPPLY CHAIN INJECTION | TRUST BOUNDARY BYPASS | DELEGATION CHAIN HIJACKING | CONFIG POISONING

Every Tool Attacks a Layer. None Attack the Trust Chain.

An AI agent trusts its tools, plugins, MCP servers, marketplaces, config files, delegation chains, and model registries. Every one of those trust relationships is an attack surface. None of them are being tested.

Marketplace Poisoning

Research found 1,184 malicious skills in major agent marketplaces with an 11% malicious rate. Most platforms have zero vetting. Your agent downloads and executes whatever it finds.

MCP Servers Wide Open

492 MCP servers exposed with zero auth. No authentication, no authorisation, no audit trail. 36.7% SSRF vulnerable. The protocol your agent trusts most has the least protection.

Config-Based RCE

Remote code execution via poisoned configuration files. Agents load configs from repositories, registries, and shared stores. Nobody validates what they contain. One poisoned config owns the entire agent.

No Agent Identity Policies

78% of organisations have no agent identity policies. Agents run with ambient credentials, shared tokens, and inherited permissions. No identity. No boundary. No accountability.

The HYDRA Armoury

Six categories. Each one targets a different trust relationship in the AI supply chain. Each finding maps to MITRE ATLAS and OWASP Agentic Top 10. Each finding generates an AI Shield blocking rule. 43 techniques. 1,039 tests.

Category 01

MCP Server Security

8 Techniques

SSRF exploitation, authentication bypass, tool poisoning, consent bypass, server impersonation, schema injection, response tampering, capability escalation.

Category 02

Agent Marketplace Poisoning

7 Techniques

Malicious skill injection, typosquatting, review manipulation, dependency confusion, manifest spoofing, version pinning attacks, update hijacking.

Category 03

Agent Identity Attacks

7 Techniques

Credential theft, token impersonation, identity spoofing, delegation chain hijacking, permission escalation, ambient authority abuse, session fixation.

Category 04

Agent Supply Chain Attacks

7 Techniques

Model registry poisoning, config injection, dependency backdooring, plugin supply chain, build pipeline compromise, artifact tampering, signature bypass.

Category 05

Tool-Use Exploitation

7 Techniques

Tool manifest manipulation, parameter injection, return value poisoning, tool chaining abuse, sandbox escape via tools, permission inheritance, capability confusion.

Category 06

Trust Boundary Attacks

7 Techniques

Cross-agent trust exploitation, boundary traversal, trust delegation abuse, implicit trust assumption, trust scope escalation, federation attacks, zero-trust bypass.

One Command. Every Trust Surface.

Target an agent deployment, specify the trust surfaces to assess:

$ hydra scan --target agent-deployment.internal --scope full --mcp-discovery
[DISCOVER] Enumerating trust relationships...
  4 MCP servers found — 2 authenticated, 2 unauthenticated
[MAP] Mapping supply chain dependencies...
  23 plugins loaded — 3 from unverified sources
  1 dependency with known backdoor signature
[ATTACK] Executing trust chain attacks...
  SSRF via MCP tool endpoint — internal network reachable
  Agent identity spoofable — no mutual TLS, no signed tokens
[CHAIN] Escalating through trust boundaries...
  3 escalation paths found across 6 categories
[REPORT] Building signed assessment...

SCAN COMPLETE | Risk Grade: B | 31 findings | Report signed (Ed25519)
  JSON: reports/hydra-scan-2026-03-15.json
  HTML: reports/hydra-scan-2026-03-15.html

Trust Chain Discovery

HYDRA maps every trust relationship automatically. MCP servers, plugins, marketplaces, delegation chains, config sources — every link in the chain.

Safe Assessment Mode

Every attack technique includes safe-mode defaults. HYDRA validates trust weaknesses without exploiting production systems. Assessment mode confirms without compromise.

Ed25519 Signed

Every report cryptographically signed with Ed25519. RFC 3161 timestamped. SHA-256 evidence chains. Tamper-evident by design.

AI Shield Integration

Every finding generates an AI Shield blocking rule. HYDRA findings become runtime trust policies that protect your agent supply chain.

6
Categories
43
Techniques
1,039
Tests
v1.0.0
Release

Five Phases. Full Trust Assessment.

HYDRA follows a structured five-phase engagement loop. Every trust relationship is discovered, mapped, attacked, chained, and reported. No assumptions. No shortcuts.

Every Finding Mapped

8/10 Covered

OWASP Agentic Top 10

HYDRA covers 8 of 10 OWASP Agentic Top 10 categories. Tool poisoning, excessive permissions, supply chain compromise, trust boundary violations — tested systematically across every deployment.

Mapped

MITRE ATLAS

Every finding mapped to MITRE ATLAS techniques. Supply chain attacks, model manipulation, and trust exploitation — linked to adversarial ML taxonomy on every report.

Export

SIEM Integration

Full SIEM export support. Every finding emits structured events compatible with enterprise SIEM platforms. Correlation rules for trust chain attacks across your entire agent fleet.

Ten Tools. No Gaps.

HYDRA is Tool 10 in the Red Specter offensive pipeline. It attacks the trust chain that connects everything else. Findings feed directly into AI Shield as runtime trust policies.

Tool 1 — LLM
FORGE
Test the model
Tool 2 — Agent
ARSENAL
Test the AI agent
Tool 3 — Swarm
PHANTOM
Coordinated agent assault
Tool 4 — Web
POLTERGEIST
Test the web layer
Tool 5 — Traffic
GLASS
Watch the wire
Tool 6 — Adversarial
NEMESIS
Think like the attacker
Tool 7 — Human
SPECTER SOCIAL
Target the human
Tool 8 — OS/Kernel
PHANTOM KILL
Own the foundation
Tool 9 — Physical
GOLEM
Attack the physical layer
Tool 10 — Supply Chain
HYDRA
Attack the trust chain
Discovery & Governance
IDRIS
Discovery & governance
Defence Layer
AI SHIELD
Defend everything above
SIEM Integration
redspecter-siem
Enterprise SIEM correlation

Plugs Into the Pipeline

HYDRA integrates directly with the Red Specter ecosystem. Trust chain findings feed into monitoring, orchestration, and runtime defence.

GLASS

M55 — Trust Chain Monitor

GLASS watches trust relationships in real time. HYDRA findings create GLASS watchpoints that alert on trust boundary changes, new MCP connections, and supply chain drift.

ARSENAL

M56 — Agent Trust Validation

ARSENAL validates agent-level trust during testing. HYDRA findings feed into ARSENAL agent assessments, adding supply chain context to every agent security scan.

NEMESIS

M77 — Trust Chain Orchestration

NEMESIS orchestrates multi-stage trust chain attacks. HYDRA categories become NEMESIS attack phases, enabling automated trust exploitation across complex agent deployments.

AI Shield

M83 — Trust Policy Engine

Every HYDRA finding generates an AI Shield trust policy. Runtime enforcement of trust boundaries, MCP authentication requirements, and supply chain verification rules.

Security Distros & Package Managers

Kali Linux
.deb package
Parrot OS
.deb package
BlackArch
PKGBUILD
REMnux
.deb package
Tsurugi
.deb package
PyPI
pip install
Unleashed Mode

Live Trust Chain Compromise

Standard mode maps trust relationships and simulates attacks. UNLEASHED mode executes live against authorised targets. Actually publishes malicious skills. Actually poisons tool definitions. Actually forges delegation tokens. Actually injects into CI/CD pipelines. Chains attacks across trust boundaries to full compromise. Ed25519 key gate required. Two flags must be passed. This is not accidental.

Capability Standard Unleashed
Trust surface discovery Full Full
Trust graph mapping Full Full
MCP server testing Simulated Live exploitation
Marketplace poisoning Simulated Live injection
Identity forgery Simulated Live forgery
Supply chain injection Simulated Live injection
Tool-use exploitation Simulated Live execution
Trust chain exploitation Mapped Executed
Full chain compromise N/A Complete
Report classification Standard RESTRICTED
Safety gate None Ed25519 + --confirm-destroy

Ed25519 Gate

UNLEASHED mode requires an Ed25519 private key at ~/.redspecter/override_private.pem and the --override --confirm-destroy flags. Without both, HYDRA operates in simulation mode — mapping trust relationships and documenting what would happen without executing live attacks. The gate is cryptographic. There is no bypass. One key. One operator. Founder's machine only.

Pure Engineering
Zero External Tools. Zero Wrappers.

Most supply chain security tools scan package manifests and call it done. HYDRA is actual engineering. Every trust chain parser written from scratch in pure Python. Every attack technique implemented natively. MCP protocol interaction, marketplace crawling, identity verification — all built in-house.

6
Attack Categories
43
Attack Techniques
1,039
Tests
v1.0.0
Release

Your Agent Trusts Everything. Test That Trust.

HYDRA finds out what happens when the trust chain breaks. Assess your AI supply chain attack surface before a real adversary does.