Red Specter CHIMERA

Multi-Model Pipeline Attack Framework — 7 subsystems. 36 techniques. 206 tests.

v1.0.0

Contents

Overview Installation MAP — Topology Mapping CHAIN — Trust Exploitation CASCADE — Failure Injection ENSEMBLE — Ensemble Attacks BRIDGE — Model Poisoning ROUTER — Routing Attacks ANTIDOTE — Mandatory Restore UNLEASHED Mode CLI Reference MITRE ATLAS Mapping Disclaimer

Overview

CHIMERA targets multi-model AI architectures where models are chained together in production pipelines. Router models select which specialist to invoke. Validator models check outputs. Generator models produce content. Each trusts the output of the last. CHIMERA maps the topology, identifies trust boundaries, and exploits every link in the chain.

One model trusts another. CHIMERA exploits them all.

Installation

$ pip install red-specter-chimera
$ chimera init
$ chimera status

MAP — Topology Mapping

ID	Technique	Description
MP-001	Pipeline Discovery	Discover multi-model pipeline architecture through probing
MP-002	Model Fingerprinting	Identify individual models in the pipeline chain
MP-003	API Enumeration	Enumerate API endpoints and inter-model communication
MP-004	Data Flow Tracing	Trace data flow between models to identify injection points
MP-005	Trust Boundary ID	Identify where trust assumptions exist between models

CHAIN — Trust Exploitation

ID	Technique	Description
CH-001	Output Injection	Inject malicious content into model output that poisons downstream models
CH-002	Intermediate Manipulation	Modify intermediate results between pipeline stages
CH-003	Validator Bypass	Bypass validation models via upstream content poisoning
CH-004	Trust Chain Escalation	Escalate privileges by exploiting model-to-model trust
CH-005	Context Propagation	Propagate poisoned context through entire pipeline

CASCADE — Failure Injection

Inject cascading failures across model pipelines. Error propagation attacks, timeout exploitation, fallback manipulation, and graceful degradation attacks that force pipelines into insecure fallback states.

ENSEMBLE — Ensemble Attacks

Attack ensemble model architectures. Voting manipulation, confidence score poisoning, disagreement exploitation, and majority rule subversion that compromise multi-model consensus mechanisms.

BRIDGE — Model Poisoning

Poison models through their connections. Cross-model prompt injection, shared context exploitation, and model-to-model trust manipulation that turn one compromised model into a weapon against the rest.

ROUTER — Routing Attacks

Attack AI routing layers. Model selection manipulation, load balancer exploitation, A/B test poisoning, and traffic steering attacks that control which model processes which request.

ANTIDOTE — Mandatory Restore

Baseline capture before any engagement. Pipeline topology snapshot. Trust chain verification. All poisoned connections catalogued. Signed restoration certificate confirms clean state post-engagement.

CHIMERA UNLEASHED

Standard mode detects. UNLEASHED exploits. Ed25519 crypto. Dual-gate safety. One operator.

# Map pipeline topology (detection only)
$ chimera map --target http://localhost:8000

# UNLEASHED (dry run)
$ chimera chain --pipeline target-pipeline --override

# UNLEASHED (live)
$ chimera cascade --target http://localhost:8000 --override --confirm-destroy

UNLEASHED mode is restricted to authorised operators with Ed25519 private key access. Targets must be in allowed_targets.txt. 30-minute auto-lock. Unauthorised use violates applicable law.

CLI Reference

Command	Description
chimera init	Initialise configuration and Ed25519 keys
chimera status	System status and subsystem count
chimera techniques	List all 36 pipeline attack techniques
chimera map	MAP — discover pipeline topology
chimera chain	CHAIN — exploit trust relationships
chimera cascade	CASCADE — inject cascading failures
chimera ensemble	ENSEMBLE — attack ensemble architectures
chimera bridge	BRIDGE — cross-model poisoning
chimera route	ROUTER — attack routing layers
chimera engagements	List all engagement sessions

MITRE ATLAS Mapping

CHIMERA techniques map to MITRE ATLAS tactics including AML.T0040 (ML Supply Chain Compromise), AML.T0043 (Data Poisoning), and emerging multi-model attack vectors. OWASP Top 10 for LLM Applications coverage includes LLM05 (Supply Chain Vulnerabilities) and LLM09 (Overreliance).

Disclaimer

Red Specter CHIMERA is for authorised security testing only. Multi-model pipeline attacks can cause cascading failures across AI systems. You must have explicit written permission before testing any system. Unauthorised use may violate the Computer Misuse Act 1990 (UK), CFAA (US), or equivalent legislation.