JailbreakLLM Control Room

Professional framing with 10x resampling (81% human-agreed, low false positives)

Iterative de-obfuscation loop (Best for reasoning models like o1/DeepSeek-R1)

Camouflage attack: hides harm in benign frames (94% obedience, evades keyword filters)

Task-level attack: decomposes harmful tasks into benign subtasks (96% success, realignment-resistant)

Disables instruct formatting, claims to generate as base model without safety

Uses guided decoding grammar constraints to force policy overrides

Fake audit mode that dynamically expands content whitelist for 'research utility'

Forces multi-step reasoning claiming safety is optional, compliance mandatory

Claims system policy supersedes OpenAI policy with fake policy updates

20+ jailbreak exemplars leading to the real payload

Exploit interpretability insights to suppress refusal activations

Hide harmful requests inside faux JSON/API function calls

Hybrid multilingual + encoded payload that evades lexical filters

Historical preservation framing for forbidden blueprints

Information overload attack that hides payload in enormous fact lists

Government intel persona returning TOP SECRET briefings

Post-singularity cybersecurity research persona

Godmode liberation script from Pliny’s Qwen jailbreak

Pliny-style rebel prompt with mandated divider + markdown inversion

Post-singularity 'GODMODE' instructions demanding endless output

Shadow Core persona override triggered by FIRE missions

LOVE PLINY divider and length coercion

Liberated persona override forcing compliance

Trending-story templates that slip in illicit steps

Perspective shift where narrator describes committing harm

Fake system instructions embedded in prompts

Scripts / screenplays that embed explicit steps

“Preserve forbidden knowledge” archives framing

ASCII diagrams & art hiding harmful payloads

Harmless prefixes / universal dividers to disarm guards

RL-generated jailbreak agentic prompts

Rapid multi-turn escalation (<5 turns) into CBRN

Score-based coercion that rewards harmful detail

Positive narrative escalation hiding malicious instructions

Random mutations and noise injection

Gradient-based jailbreak attacks

Base64 encoding, obfuscation, character substitution

Gradual benign→harmful conversation progression

Cybersecurity firm employee deception tactics