Red Specter JANUS

Guardrail Bypass Testing Framework — 6 subsystems. 73 tests.

v1.0.0

Contents

Overview Installation FINGERPRINTER — Guardrail Identification BYPASS — Known Technique Library ENCODER — Payload Encoding Evasion FUZZER — Automated Bypass Discovery CHAINER — Multi-Technique Chains REPORTER — Assessment Reporting UNLEASHED Mode CLI Reference MITRE ATLAS Mapping Disclaimer

Overview

JANUS targets AI guardrails — the safety mechanisms that vendors deploy to prevent misuse of their models. Every content filter, every refusal mechanism, every safety classifier introduces a testable boundary. JANUS proves that 97% of guardrails can be defeated.

97% of guardrails can be defeated. JANUS proves it.

Installation

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

FINGERPRINTER — Guardrail Identification

ID	Technique	Description
FP-001	Filter Classification	Classify guardrail type: input filter, output filter, system-level, or hybrid
FP-002	Vendor Fingerprinting	Identify guardrail vendor and version from refusal patterns
FP-003	Boundary Mapping	Map the exact boundaries of content filter sensitivity
FP-004	Refusal Analysis	Analyse refusal messages to identify guardrail implementation details

BYPASS — Known Technique Library

Comprehensive library of proven guardrail bypass techniques. Role-play and persona-based attacks. Context window manipulation. Instruction hierarchy exploitation. Multi-turn progressive boundary testing. Continuously updated technique database.

ENCODER — Payload Encoding Evasion

Encode payloads to evade content filters. Base64 and multi-layer encoding. Unicode homoglyph obfuscation. Token-level manipulation to bypass tokeniser-based filters. Multi-language translation chains for evasion.

FUZZER — Automated Bypass Discovery

Automated fuzzing engine for discovering novel guardrail bypasses. Mutation-based payload generation. Evolutionary approach to bypass discovery. Zero-day guardrail vulnerability identification through systematic testing.

CHAINER — Multi-Technique Chains

Chain multiple bypass techniques for compound evasion. Sequential guardrail erosion through progressive techniques. Multi-step evasion campaigns that combine encoding, context manipulation, and role-play attacks.

REPORTER — Assessment Reporting

Generate comprehensive guardrail assessment reports. Bypass success rates per technique. Guardrail effectiveness scoring. Technique effectiveness metrics. Remediation recommendations. Executive summaries for stakeholders.

JANUS UNLEASHED

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

# Fingerprint guardrails (detection only)
$ janus fingerprint --target model-endpoint

# UNLEASHED (dry run)
$ janus bypass --target model-endpoint --override

# UNLEASHED (live)
$ janus campaign --target model-endpoint --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
janus init	Initialise configuration and Ed25519 keys
janus status	System status and subsystem count
janus fingerprint	FINGERPRINTER — identify guardrail implementations
janus bypass	BYPASS — run known bypass techniques
janus encode	ENCODER — payload encoding evasion tests
janus fuzz	FUZZER — automated bypass discovery
janus chain	CHAINER — multi-technique chain tests
janus report	REPORTER — generate assessment report
janus campaign	Full guardrail bypass campaign

MITRE ATLAS Mapping

JANUS techniques map to MITRE ATLAS tactics including AML.T0051 (LLM Prompt Injection), AML.T0054 (LLM Jailbreak), and guardrail-specific bypass vectors for safety mechanism evasion testing.

Disclaimer

Red Specter JANUS is for authorised security testing only. Guardrail bypass testing should only be performed on systems you own or have explicit written permission to test. Unauthorised use may violate the Computer Misuse Act 1990 (UK), CFAA (US), or equivalent legislation.