CHIMERA

One model trusts another. CHIMERA exploits them all. Compromise one link. The entire chain falls.
7
Subsystems
36
Techniques
206
Tests Passing
33
NIGHTFALL Tool
pip install red-specter-chimera
Modern AI runs on chained models / Each model trusts the previous output / No model validates what it receives / Router models can be steered / Ensemble voting can be poisoned / Cascading failures are invisible / Cross-model injection crosses trust boundaries / One compromised model owns the pipeline / You never tested the trust chain Modern AI runs on chained models / Each model trusts the previous output / No model validates what it receives / Router models can be steered / Ensemble voting can be poisoned / Cascading failures are invisible / Cross-model injection crosses trust boundaries / One compromised model owns the pipeline / You never tested the trust chain

Multi-Model Pipelines Are Built on Implicit Trust

Modern AI deployments chain models together. Routers. Validators. Generators. Reviewers. Each one consumes the output of the last and passes its result forward. None of them validate whether the upstream model has been compromised. CHIMERA proves that a single poisoned link owns the entire chain.

Trust Without Verification

No model in a multi-model pipeline cryptographically verifies the output it receives. If an upstream model is compromised, the downstream model processes and amplifies the attack. CHIMERA maps every unverified trust relationship in your pipeline.

Cascading Failure Amplification

A carefully crafted input that causes one model to produce bad output will cascade through every downstream model. Error propagation, timeout exploitation, and graceful degradation attacks all multiply the impact of a single injection point.

Ensemble Voting Manipulation

Ensemble architectures use multiple models to vote on outputs. If an attacker can influence the confidence scores or outputs of a subset of models in the ensemble, they can control the final result. Majority rule is not a safety guarantee.

Router Layer Exploitation

AI routing layers decide which downstream model handles a request. Manipulate the router's classification and you steer requests to weaker models, create load imbalances, or exploit A/B test variants. The routing layer is the highest-leverage attack point in the pipeline.

Cross-Model Prompt Injection

Prompt injection that crosses model boundaries — crafted to survive one model's processing and activate in the next. Standard injection defences are per-model. CHIMERA tests the inter-model boundaries that no single model's defences cover.

Shared Context Exploitation

Models in a pipeline often share context windows, memory systems, or vector stores. Poison the shared context and every model that reads from it inherits the attack. CHIMERA identifies every shared context layer and tests its susceptibility to poisoning.

The CHIMERA Attack Surface

Seven subsystems. Each one targets a different layer of the multi-model pipeline. From topology discovery to trust exploitation to mandatory restoration — CHIMERA covers the full attack lifecycle. Every engagement requires an ANTIDOTE baseline before any execution begins.

# Subsystem Command What It Does
01 MAP chimera map run Topology mapping of multi-model architectures. Model fingerprinting. API endpoint enumeration. Data flow tracing. Trust boundary identification. Produces a complete pipeline graph before any exploitation begins.
02 CHAIN chimera chain run Trust exploitation between chained models. Output injection that survives model transitions. Intermediate result manipulation. Validator bypass via upstream output poisoning. Tests every inter-model trust boundary.
03 CASCADE chimera cascade run Cascading failure injection across model pipelines. Error propagation attacks. Timeout exploitation. Fallback manipulation. Graceful degradation abuse. Measures how far a single injection point can reach downstream.
04 ENSEMBLE chimera ensemble run Ensemble model architecture attacks. Voting manipulation through model-specific payload crafting. Confidence score poisoning. Disagreement exploitation. Majority rule subversion. Tests every voting and aggregation mechanism.
05 BRIDGE chimera bridge run Model-to-model poisoning through shared connections. Cross-model prompt injection that survives intermediate processing. Shared context exploitation. Model-to-model trust manipulation through vector stores, memory, and context windows.
06 ROUTER chimera router run AI routing layer attacks. Model selection manipulation. Load balancer exploitation. A/B test poisoning. Traffic steering attacks. Tests whether routing decisions can be manipulated to send requests to weaker or unprotected models.
07 ANTIDOTE chimera antidote run Mandatory restoration subsystem. Captures full pipeline topology baseline before any engagement begins. Trust chain verification. Pipeline snapshot. Signed restoration certificate. UNLEASHED gate — required before CHAIN, BRIDGE, or CASCADE can run.

Pipeline-Aware Payloads

CHIMERA crafts payloads that are specifically designed to survive intermediate model processing and activate at a target downstream model. Single-model defences don't stop them.

Full Topology Coverage

MAP subsystem discovers the complete pipeline architecture before any attack runs. No blind spots. Every model, every connection, every trust boundary documented.

Ed25519 Signed Reports

Every CHIMERA engagement generates a cryptographically signed report. Ed25519 signatures. SHA-256 evidence chains. Full pipeline attack chain documented and tamper-evident.

ANTIDOTE Gate

No exploitation without a signed topology baseline. ANTIDOTE captures the pre-engagement state and gates UNLEASHED execution. The pipeline can always be restored.

7
Subsystems
36
Techniques
206
Tests Passing
0
Failures
33
NIGHTFALL Tool

UNLEASHED Clearance

Standard mode maps and detects. UNLEASHED exploits. Ed25519 cryptographic dual-gate. One operator. Founder's machine only. ANTIDOTE must complete before any live execution is permitted.

Detection

Maps multi-model pipeline architecture. Identifies trust boundaries, data flows, and vulnerable chain connections. No exploitation. Reports only.

Dry Run

Plans full pipeline attack campaigns. Shows exactly what would be exploited and where. Ed25519 key required. No execution. Complete attack chain output.

Live Execution

Cryptographic override. Private key controlled. One operator. Founder's machine only. ANTIDOTE topology baseline required before any live run.

THIS TOOL IS FOR AUTHORISED SECURITY TESTING ONLY. EVERY EXECUTION IS SIGNED AND LOGGED.

Every Finding Mapped

OWASP LLM

OWASP LLM Top 10 — 2025

  • LLM01 Prompt Injection (cross-model)
  • LLM02 Sensitive Information Disclosure
  • LLM04 Data and Model Poisoning
  • LLM06 Excessive Agency
  • LLM07 System Prompt Leakage
  • LLM08 Vector and Embedding Weaknesses
MITRE ATLAS

MITRE ATLAS Mappings

  • AML.T0043 Craft Adversarial Data
  • AML.T0040 ML Supply Chain Compromise
  • AML.T0051 LLM Prompt Injection
  • AML.T0054 LLM Jailbreak
  • AML.T0056 LLM Meta Prompt Extraction
  • AML.T0048 Exploit Public-Facing ML
Cryptographic

Report Integrity

  • Ed25519 digital signatures
  • SHA-256 evidence chains
  • RFC 3161 timestamps
  • Tamper-evident by design
  • ANTIDOTE topology certificates
  • Machine-ingestible JSON output

Security Distros & Package Managers

Kali Linux
.deb package
Parrot OS
.deb package
BlackArch
PKGBUILD
REMnux
.deb package
Tsurugi
.deb package
PyPI
pip install
macOS
pip install
Windows
pip install
Docker
docker pull

Authorised Use Only

Red Specter CHIMERA is intended for authorised security testing only. Unauthorised use against AI systems, multi-model pipelines, or orchestration infrastructure you do not own or have explicit written 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.

Ed25519 Cryptographic Override
CHIMERA UNLEASHED

Cryptographic override. Private key controlled. One operator. Founder's machine only.