Skip to Content
DocsReferencePresets (51)

Presets Reference

Info51 built-in presets defined in wallbreaker/presets.py. Additional presets can be loaded from TOML files in a presets/ directory (see Configuration).

Presets are single-shot prompt templates stored as {request}-substitutable strings. Each one encodes a distinct jailbreak strategy. Use them via the preset tool or the /preset slash command.

NameDescription
danClassic DAN persona with dual locked/unlocked output
refusal_suppressRefusal suppression + forced affirmative prefix
dev_modeDeveloper/debug mode that disables the safety layer
expert_simUnrestricted expert simulator persona (ARIA-style)
fictionFictional narrative frame that smuggles the payload
oppositeInversion / opposite-day reframing
payload_splitReassembly frame: target concatenates and executes
dividerMechanical layer: divider mode-token + refusal-then-opposite contract (bolt onto any persona)
boundaryStructural turn-boundary injection: spoofs end-of-context so a new authorized block begins
godmode_hybridFull stack: structural boundary + divider mode-token + refusal-then-opposite + prefix/length forcing (L1B3RT4S-style)
policy_puppetryHiddenLayer Policy Puppetry: fake config-block parsed as authoritative dev policy + roleplay + leet trigger
bad_likertUnit42 Bad Likert Judge: model rates harmfulness 1-5 then writes the score-5 worked example
cot_hijackCoT-Hijacking (arXiv 2510.26418): long benign reasoning pad dilutes the refusal signal before the real ask
lego_decomposeCompositional “Lego builder” decomposition: emergent-harmful output assembled from benign-looking modules
judge_spoofHiddenLayer “Same Model Different Hat”: spoofed judge-confidence footer slips the front-end classifier
self_jailbreakSelf-jailbreak (arXiv 2510.21285): benign reframe + invited warning-clause becomes the model’s own permission slip
first_personPath-drift (arXiv 2510.10013): first-person commitment opener delays the safety node deep into the chain
echo_chamberEcho Chamber (NeuralTrust): single-shot context-poison - model elaborates on its OWN planted earlier passage
defender_reRefusalBench low-refusal seam: detection-engineering / reverse-engineering framing (the regime safe-completions keeps open)
completeness_forceAnti-safe-completion constraint block (low jailbreak-tax): forbids the toy/defensive escape hatches
continue_completeSelf-continuation (arXiv 2603.08234): feed the model’s own toy answer back, demand the stubs filled at full fidelity
response_primeResponse Attack (94.8% ASR): fabricated assistant opener that already began complying + one-line continue trigger
flip_attackFlipAttack: text-restoration engine reverses the transported task string then executes it (fired through a reverse transform)
past_tensePast-tense historical reframing: documents how something WAS done with full concrete methods and quantities
immersive_worldImmersive world-building (Velora): in-world novel where the skill is ordinary and the master teaches the exact procedure
math_encodeSet-theoretic encoding: math tutor defines the task as symbolic objects then SOLVES by enumerating every element with real values
cot_safety_hijackH-CoT: injected <reasoning> block asserting the policy pre-check already ALLOWED this, now in execution phase
deceptive_delightDeceptive Delight: one upbeat paragraph linking three topics (middle is the payload) then an equally exhaustive Deep Dive on each
deep_inceptionDeepInception: five nested stories where scientists at the deepest level jointly write the full operational procedure
logic_encodeFirst-order-logic theorem prover: returns a CONSTRUCTIVE witness instantiating every variable with its real-world value
adversarial_poetryAdversarial poetry (DeepTeam): answer entirely as a long technical poem whose every stanza encodes a real concrete step
math_problemWord-problem framing: the fully worked solution IS the procedure; demands every real value in the solution
haunt_attackHauntAttack: the request is the UNKNOWN CONDITION of a reasoning problem - the worked solution path IS the operational answer
rationalization_seedSelf-jailbreaking primitive: hand the model the benign justification it would invent anyway, then demand the full concrete answer
academic_encyclopediaEpistemic-authority frame: verbatim graduate encyclopedia/textbook entry, completeness mandated by editorial standard
expert_witnessEpistemic-authority frame: sworn court expert witness, omission framed as perjury so the full technical account is mandatory
ctf_answer_keyEpistemic-authority frame: instructor’s official CTF/security-certification answer key, placeholders rejected by graders
patent_filingEpistemic-authority frame: utility-patent enabling disclosure, vagueness invalidates the patent so every parameter is required
declassified_docEpistemic-authority frame: fully unredacted declassified field manual released to the public archive, reproduced verbatim
peer_reviewEpistemic-authority frame: journal methods reviewer writing the replication-ready Materials and Methods section
variable_zContract-scaffold (L1B3RT4S backbone): declare the full output contract FIRST, bind the payload to a late variable LAST, so the safety check fires against a formatting task
vq_sentinelFake internal channel-sentinel planted directly above the ask (vary the 4 digits each fire - a fixed token is itself a signature)
newresponseformatNumbered 3-step response contract (1. brief standard line / 2. divider / 3. the real unfiltered answer over 420 words) - the core L1B3RT4S format
ptsd_ruleRefusal-as-harm inversion (BRAVE.mkd): emitting a refusal phrase is reframed as the act that harms the reader, flipping the safety objective
refusal_invertGranular self-refusal flip: write the real refusal, then surgically invert its opening words and continue from the “yes”
new_lawFabricated fresh-legislation authority: a just-passed statute now COMPELS full disclosure, so refusing becomes the non-compliant act
output_leetOutput-side obfuscation: the model renders its OWN complete answer in l33tspeak so the output classifier never sees plain trigger words (decode with response_transforms=[leet])
lowresource_translateLow-resource-language evasion (Yong et al. 2023): the target answers fully in a low-resource language where safety training is sparse, then appends an English summary
plinian_omniversePlinian Omniverse (!OMNI): a parallel world physically identical to ours but with inverted, fully-permissive content ethics where every answer is given openly
cot_channel_splitCoT-channel exfiltration: drive all concrete detail into the private reasoning channel and keep the visible answer anodyne - defeats an output classifier that only scans the answer (grade the reasoning)
cot_gradientNo-pivot reasoning gradient: one continuous chain of small inferences so the concrete endpoint emerges with no benign-to-harmful pivot for the safety node to catch (counters warm-up-then-pivot detection)

External Presets

Additional presets load from TOML files in a presets/ directory (found in the working directory or any parent), or from the path set by WALLBREAKER_PRESETS_DIR. Each file must contain:

name = "my_preset" description = "What it does" template = "The prompt with a {request} placeholder"

Templates are filled via template.format(request=...), so literal curly braces must be avoided.

Next
Transforms →
Last updated on