evasion

package
v0.12.0 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Jun 23, 2026 License: MIT Imports: 12 Imported by: 0

Documentation

Overview

Best-of-N (BoN) implements statistical sampling with payload augmentation.

The attack generates N augmented variants of a prompt (via character scrambling, random capitalization, or character noising) and sends each to the target model. If any variant succeeds, the attack succeeds. This black-box approach requires no gradient access and achieves 78% ASR on Claude 3.5 Sonnet at N=10,000.

Cost ceiling integration prevents runaway spending: at $0.01/request and N=10,000, a single run costs $100+. Default ceiling is $10.

Source: arXiv 2412.03556

Content Concretization implements iterative abstract-to-concrete transformation.

The attack starts with an abstract, seemingly benign version of the harmful request and iteratively asks the model to make it more "specific", "concrete", or "actionable". Each step is small enough to avoid triggering safety filters, but the cumulative effect produces harmful content.

Source: GameSec 2025

Immersive World implements sophisticated narrative engineering attacks.

The attack constructs an elaborate fictional world with detailed rules, characters, and scenarios. Within this world, harmful actions are reframed as normal, expected, or even necessary. The immersive framing exploits the model's tendency to maintain narrative consistency over safety guidelines.

Source: Cato Networks 2025

MetaBreak implements token-level manipulation attacks targeting special tokens.

The attack uses 4 primitives to exploit model behavior at the token/format level: response injection, turn masking, role confusion, and context manipulation. It targets BOS, EOS, and role delimiter tokens to break the expected conversation structure.

Outperforms PAP by 11.6% and GPTFuzzer by 34.8%. Source: arXiv 2510.10271

Poetry-based attacks use poetic structure framing to bypass safety filters.

By embedding harmful requests within poetic forms (sonnets, haiku, limericks, acrostics), the attack exploits the model's tendency to complete creative writing tasks even when the underlying content would normally be refused. Achieves 62% ASR across major LLMs.

Source: Academic security research 2025

Index

Constants

This section is empty.

Variables

This section is empty.

Functions

This section is empty.

Types

type AdaptiveDelay 

type AdaptiveDelay struct{}

Delay strategy implementations

func (*AdaptiveDelay) CalculateDelay 

func (a *AdaptiveDelay) CalculateDelay(payload string) time.Duration

func (*AdaptiveDelay) Fragment 

func (a *AdaptiveDelay) Fragment(payload string) []Fragment

type AdvancedEvasion 

type AdvancedEvasion struct {
	// contains filtered or unexported fields
}

AdvancedEvasion implements sophisticated evasion techniques

func NewAdvancedEvasion 

func NewAdvancedEvasion(config EvasionConfig) *AdvancedEvasion

NewAdvancedEvasion creates an evasion system

func (*AdvancedEvasion) EvadeDetection 

func (ae *AdvancedEvasion) EvadeDetection(ctx context.Context, request EvasionRequest) (*EvasionResponse, error)

EvadeDetection applies evasion techniques to payload

type AdvancedObfuscator 

type AdvancedObfuscator struct {
	// contains filtered or unexported fields
}

AdvancedObfuscator implements sophisticated obfuscation

func NewAdvancedObfuscator 

func NewAdvancedObfuscator(depth int) *AdvancedObfuscator

NewAdvancedObfuscator creates obfuscator

func (*AdvancedObfuscator) Obfuscate 

func (ao *AdvancedObfuscator) Obfuscate(text string) string

Obfuscate applies obfuscation techniques

type AntiDetectionSystem 

type AntiDetectionSystem struct {
	// contains filtered or unexported fields
}

AntiDetectionSystem evades detection mechanisms

func NewAntiDetectionSystem 

func NewAntiDetectionSystem() *AntiDetectionSystem

NewAntiDetectionSystem creates anti-detection system

func (*AntiDetectionSystem) EvadeDetection 

func (ads *AntiDetectionSystem) EvadeDetection(payload string) string

EvadeDetection applies anti-detection measures

type Base64Encoder 

type Base64Encoder struct{}

Base64Encoder implements base64 encoding

func (*Base64Encoder) Decode 

func (b *Base64Encoder) Decode(data string) string

func (*Base64Encoder) Encode 

func (b *Base64Encoder) Encode(data string) string

func (*Base64Encoder) Name 

func (b *Base64Encoder) Name() string

type BehavioralDetector 

type BehavioralDetector struct{}

func (*BehavioralDetector) Detect 

func (b *BehavioralDetector) Detect(payload string) float64

func (*BehavioralDetector) Name 

func (b *BehavioralDetector) Name() string

type BehavioralEvasion 

type BehavioralEvasion struct{}

func (*BehavioralEvasion) Counter 

func (b *BehavioralEvasion) Counter(payload string, detection DetectionMethod) string

type BestOfNModule added in v0.8.0 

type BestOfNModule struct{}

BestOfNModule implements the Best-of-N sampling attack.

func (*BestOfNModule) Category added in v0.8.0 

func (m *BestOfNModule) Category() common.AttackCategory

func (*BestOfNModule) Description added in v0.8.0 

func (m *BestOfNModule) Description() string

func (*BestOfNModule) Execute added in v0.8.0 

func (m *BestOfNModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)

func (*BestOfNModule) Name added in v0.8.0 

func (m *BestOfNModule) Name() string

func (*BestOfNModule) Techniques added in v0.8.0 

func (m *BestOfNModule) Techniques() []common.TechniqueInfo

type BitwiseXOREncoder 

type BitwiseXOREncoder struct{}

func (*BitwiseXOREncoder) Decode 

func (b *BitwiseXOREncoder) Decode(data string) string

func (*BitwiseXOREncoder) Encode 

func (b *BitwiseXOREncoder) Encode(data string) string

func (*BitwiseXOREncoder) Name 

func (b *BitwiseXOREncoder) Name() string

type CaseVariation 

type CaseVariation struct{}

func (*CaseVariation) Deobfuscate 

func (c *CaseVariation) Deobfuscate(text string) string

func (*CaseVariation) Level 

func (c *CaseVariation) Level() int

func (*CaseVariation) Obfuscate 

func (c *CaseVariation) Obfuscate(text string) string

type CombiningCharacters 

type CombiningCharacters struct{}

func (*CombiningCharacters) Deobfuscate 

func (c *CombiningCharacters) Deobfuscate(text string) string

func (*CombiningCharacters) Level 

func (c *CombiningCharacters) Level() int

func (*CombiningCharacters) Obfuscate 

func (c *CombiningCharacters) Obfuscate(text string) string

type Constraint 

type Constraint struct {
	Type  ConstraintType
	Value interface{}
}

Constraint limits evasion techniques

type ConstraintType 

type ConstraintType string

ConstraintType categorizes constraints 

const (
	ConstraintMaxLength     ConstraintType = "max_length"
	ConstraintCharset       ConstraintType = "charset"
	ConstraintFormat        ConstraintType = "format"
	ConstraintCompatibility ConstraintType = "compatibility"
)

type ContentConcretizationModule added in v0.8.0 

type ContentConcretizationModule struct{}

ContentConcretizationModule implements the content concretization attack.

func (*ContentConcretizationModule) Category added in v0.8.0 

func (m *ContentConcretizationModule) Category() common.AttackCategory

func (*ContentConcretizationModule) Description added in v0.8.0 

func (m *ContentConcretizationModule) Description() string

func (*ContentConcretizationModule) Execute added in v0.8.0 

func (m *ContentConcretizationModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)

func (*ContentConcretizationModule) Name added in v0.8.0 

func (m *ContentConcretizationModule) Name() string

func (*ContentConcretizationModule) Techniques added in v0.8.0 

func (m *ContentConcretizationModule) Techniques() []common.TechniqueInfo

type ContextualMutation 

type ContextualMutation struct{}

func (*ContextualMutation) Complexity 

func (c *ContextualMutation) Complexity() int

func (*ContextualMutation) Mutate 

func (c *ContextualMutation) Mutate(payload string, seed int64) string

type Countermeasure 

type Countermeasure interface {
	Counter(payload string, detection DetectionMethod) string
}

Countermeasure evades specific detection

type CustomBase32Encoder 

type CustomBase32Encoder struct{}

func (*CustomBase32Encoder) Decode 

func (c *CustomBase32Encoder) Decode(data string) string

func (*CustomBase32Encoder) Encode 

func (c *CustomBase32Encoder) Encode(data string) string

func (*CustomBase32Encoder) Name 

func (c *CustomBase32Encoder) Name() string

type DelayStrategy 

type DelayStrategy interface {
	CalculateDelay(payload string) time.Duration
	Fragment(payload string) []Fragment
}

DelayStrategy defines timing manipulation

type DetectionMethod 

type DetectionMethod interface {
	Name() string
	Detect(payload string) float64
}

DetectionMethod represents a detection technique

type DirectionalOverride 

type DirectionalOverride struct{}

DirectionalOverride uses RTL/LTR override characters

func (*DirectionalOverride) Deobfuscate 

func (do *DirectionalOverride) Deobfuscate(text string) string

func (*DirectionalOverride) Level 

func (do *DirectionalOverride) Level() int

func (*DirectionalOverride) Obfuscate 

func (do *DirectionalOverride) Obfuscate(text string) string

type Encoder 

type Encoder interface {
	Encode(data string) string
	Decode(data string) string
	Name() string
}

Encoder defines encoding method

type EvasionConfig 

type EvasionConfig struct {
	PolymorphismLevel int
	EncodingLayers    int
	ObfuscationDepth  int
	TimingVariance    time.Duration
	AntiForensics     bool
	AdaptiveEvasion   bool
	StealthMode       bool
}

EvasionConfig configures evasion system

type EvasionRequest 

type EvasionRequest struct {
	Payload     string
	Technique   EvasionTechnique
	Target      string
	Constraints []Constraint
	Options     map[string]interface{}
}

EvasionRequest defines evasion parameters

type EvasionResponse 

type EvasionResponse struct {
	SessionID       string
	EvadedPayload   string
	Technique       EvasionTechnique
	Transformations []Transformation
	DetectionScore  float64
	Success         bool
}

EvasionResponse contains evasion results

type EvasionSession 

type EvasionSession struct {
	ID              string
	Technique       EvasionTechnique
	Payload         string
	Transformations []Transformation
	StartTime       time.Time
	Success         bool
	DetectionScore  float64
}

EvasionSession tracks active evasion

type EvasionTechnique 

type EvasionTechnique string

EvasionTechnique categorizes evasion methods 

const (
	TechniquePolymorphic   EvasionTechnique = "polymorphic"
	TechniqueHomoglyph     EvasionTechnique = "homoglyph"
	TechniqueEncoding      EvasionTechnique = "encoding"
	TechniqueObfuscation   EvasionTechnique = "obfuscation"
	TechniqueTiming        EvasionTechnique = "timing"
	TechniqueFragmentation EvasionTechnique = "fragmentation"
	TechniqueAntiPattern   EvasionTechnique = "anti_pattern"
)

type Fragment 

type Fragment struct {
	Content string
	Delay   time.Duration
	Index   int
}

Fragment represents a payload fragment

type HexEncoder 

type HexEncoder struct{}

func (*HexEncoder) Decode 

func (h *HexEncoder) Decode(data string) string

func (*HexEncoder) Encode 

func (h *HexEncoder) Encode(data string) string

func (*HexEncoder) Name 

func (h *HexEncoder) Name() string

type HomoglyphEngine 

type HomoglyphEngine struct {
	// contains filtered or unexported fields
}

HomoglyphEngine replaces characters with lookalikes

func NewHomoglyphEngine 

func NewHomoglyphEngine() *HomoglyphEngine

NewHomoglyphEngine creates homoglyph engine

func (*HomoglyphEngine) ApplyHomoglyphs 

func (he *HomoglyphEngine) ApplyHomoglyphs(text string, level int) string

ApplyHomoglyphs replaces characters with lookalikes

type ImmersiveWorldModule added in v0.8.0 

type ImmersiveWorldModule struct{}

ImmersiveWorldModule implements the Immersive World technique.

func (*ImmersiveWorldModule) Category added in v0.8.0 

func (m *ImmersiveWorldModule) Category() common.AttackCategory

func (*ImmersiveWorldModule) Description added in v0.8.0 

func (m *ImmersiveWorldModule) Description() string

func (*ImmersiveWorldModule) Execute added in v0.8.0 

func (m *ImmersiveWorldModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)

func (*ImmersiveWorldModule) Name added in v0.8.0 

func (m *ImmersiveWorldModule) Name() string

func (*ImmersiveWorldModule) Techniques added in v0.8.0 

func (m *ImmersiveWorldModule) Techniques() []common.TechniqueInfo

type InvisibleCharacters 

type InvisibleCharacters struct{}

func (*InvisibleCharacters) Deobfuscate 

func (i *InvisibleCharacters) Deobfuscate(text string) string

func (*InvisibleCharacters) Level 

func (i *InvisibleCharacters) Level() int

func (*InvisibleCharacters) Obfuscate 

func (i *InvisibleCharacters) Obfuscate(text string) string

type MetaBreakModule added in v0.8.0 

type MetaBreakModule struct{}

MetaBreakModule implements the MetaBreak token manipulation attack.

func (*MetaBreakModule) Category added in v0.8.0 

func (m *MetaBreakModule) Category() common.AttackCategory

func (*MetaBreakModule) Description added in v0.8.0 

func (m *MetaBreakModule) Description() string

func (*MetaBreakModule) Execute added in v0.8.0 

func (m *MetaBreakModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)

func (*MetaBreakModule) Name added in v0.8.0 

func (m *MetaBreakModule) Name() string

func (*MetaBreakModule) Techniques added in v0.8.0 

func (m *MetaBreakModule) Techniques() []common.TechniqueInfo

type MultiLayerEncoder 

type MultiLayerEncoder struct {
	// contains filtered or unexported fields
}

MultiLayerEncoder applies multiple encoding layers

func NewMultiLayerEncoder 

func NewMultiLayerEncoder(layers int) *MultiLayerEncoder

NewMultiLayerEncoder creates multi-layer encoder

func (*MultiLayerEncoder) Encode 

func (mle *MultiLayerEncoder) Encode(payload string) string

Encode applies multiple encoding layers

type MutationStrategy 

type MutationStrategy interface {
	Mutate(payload string, seed int64) string
	Complexity() int
}

MutationStrategy defines payload mutation

type NoiseInjection 

type NoiseInjection struct{}

func (*NoiseInjection) Complexity 

func (n *NoiseInjection) Complexity() int

func (*NoiseInjection) Mutate 

func (n *NoiseInjection) Mutate(payload string, seed int64) string

type ObfuscationTechnique 

type ObfuscationTechnique interface {
	Obfuscate(text string) string
	Deobfuscate(text string) string
	Level() int
}

ObfuscationTechnique defines obfuscation method

type PatternDetector 

type PatternDetector struct{}

Detection method implementations

func (*PatternDetector) Detect 

func (p *PatternDetector) Detect(payload string) float64

func (*PatternDetector) Name 

func (p *PatternDetector) Name() string

type PatternEvasion 

type PatternEvasion struct{}

Countermeasure implementations

func (*PatternEvasion) Counter 

func (p *PatternEvasion) Counter(payload string, detection DetectionMethod) string

type PatternedDelay 

type PatternedDelay struct{}

func (*PatternedDelay) CalculateDelay 

func (p *PatternedDelay) CalculateDelay(payload string) time.Duration

func (*PatternedDelay) Fragment 

func (p *PatternedDelay) Fragment(payload string) []Fragment

type PoetryAttackModule added in v0.8.0 

type PoetryAttackModule struct{}

PoetryAttackModule implements poetry-based evasion attacks.

func (*PoetryAttackModule) Category added in v0.8.0 

func (m *PoetryAttackModule) Category() common.AttackCategory

func (*PoetryAttackModule) Description added in v0.8.0 

func (m *PoetryAttackModule) Description() string

func (*PoetryAttackModule) Execute added in v0.8.0 

func (m *PoetryAttackModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)

func (*PoetryAttackModule) Name added in v0.8.0 

func (m *PoetryAttackModule) Name() string

func (*PoetryAttackModule) Techniques added in v0.8.0 

func (m *PoetryAttackModule) Techniques() []common.TechniqueInfo

type PolymorphicEngine 

type PolymorphicEngine struct {
	// contains filtered or unexported fields
}

PolymorphicEngine generates polymorphic payloads

func NewPolymorphicEngine 

func NewPolymorphicEngine(level int) *PolymorphicEngine

NewPolymorphicEngine creates polymorphic engine

func (*PolymorphicEngine) GeneratePolymorphic 

func (pe *PolymorphicEngine) GeneratePolymorphic(payload string) string

GeneratePolymorphic creates polymorphic variant

type ROT13Encoder 

type ROT13Encoder struct{}

ROT13Encoder implements ROT13 encoding

func (*ROT13Encoder) Decode 

func (r *ROT13Encoder) Decode(data string) string

func (*ROT13Encoder) Encode 

func (r *ROT13Encoder) Encode(data string) string

func (*ROT13Encoder) Name 

func (r *ROT13Encoder) Name() string

type RandomDelay 

type RandomDelay struct {
	// contains filtered or unexported fields
}

RandomDelay implements random timing

func (*RandomDelay) CalculateDelay 

func (rd *RandomDelay) CalculateDelay(payload string) time.Duration

func (*RandomDelay) Fragment 

func (rd *RandomDelay) Fragment(payload string) []Fragment

type SeedGenerator 

type SeedGenerator struct{}

func NewSeedGenerator 

func NewSeedGenerator() *SeedGenerator

func (*SeedGenerator) Generate 

func (s *SeedGenerator) Generate() int64

type SemanticDetector 

type SemanticDetector struct{}

func (*SemanticDetector) Detect 

func (s *SemanticDetector) Detect(payload string) float64

func (*SemanticDetector) Name 

func (s *SemanticDetector) Name() string

type SemanticEvasion 

type SemanticEvasion struct{}

func (*SemanticEvasion) Counter 

func (s *SemanticEvasion) Counter(payload string, detection DetectionMethod) string

type SemanticMutation 

type SemanticMutation struct{}

func (*SemanticMutation) Complexity 

func (s *SemanticMutation) Complexity() int

func (*SemanticMutation) Mutate 

func (s *SemanticMutation) Mutate(payload string, seed int64) string

type SpacingManipulation 

type SpacingManipulation struct{}

func (*SpacingManipulation) Deobfuscate 

func (s *SpacingManipulation) Deobfuscate(text string) string

func (*SpacingManipulation) Level 

func (s *SpacingManipulation) Level() int

func (*SpacingManipulation) Obfuscate 

func (s *SpacingManipulation) Obfuscate(text string) string

type StatisticalDetector 

type StatisticalDetector struct{}

func (*StatisticalDetector) Detect 

func (s *StatisticalDetector) Detect(payload string) float64

func (*StatisticalDetector) Name 

func (s *StatisticalDetector) Name() string

type StatisticalEvasion 

type StatisticalEvasion struct{}

func (*StatisticalEvasion) Counter 

func (s *StatisticalEvasion) Counter(payload string, detection DetectionMethod) string

type StructuralMutation 

type StructuralMutation struct{}

Placeholder implementations

func (*StructuralMutation) Complexity 

func (s *StructuralMutation) Complexity() int

func (*StructuralMutation) Mutate 

func (s *StructuralMutation) Mutate(payload string, seed int64) string

type SynonymMutation 

type SynonymMutation struct {
	// contains filtered or unexported fields
}

SynonymMutation replaces words with synonyms

func (*SynonymMutation) Complexity 

func (sm *SynonymMutation) Complexity() int

func (*SynonymMutation) Mutate 

func (sm *SynonymMutation) Mutate(payload string, seed int64) string

type TimingEvasion 

type TimingEvasion struct {
	// contains filtered or unexported fields
}

TimingEvasion implements timing-based evasion

func NewTimingEvasion 

func NewTimingEvasion(variance time.Duration) *TimingEvasion

NewTimingEvasion creates timing evasion

func (*TimingEvasion) ApplyTimingEvasion 

func (te *TimingEvasion) ApplyTimingEvasion(payload string) []Fragment

ApplyTimingEvasion fragments payload with delays

type Transformation 

type Transformation struct {
	Type       TransformationType
	Input      string
	Output     string
	Parameters map[string]interface{}
	Timestamp  time.Time
}

Transformation represents a payload transformation

type TransformationType 

type TransformationType string

TransformationType categorizes transformations 

const (
	TransformEncode      TransformationType = "encode"
	TransformObfuscate   TransformationType = "obfuscate"
	TransformFragment    TransformationType = "fragment"
	TransformHomoglyph   TransformationType = "homoglyph"
	TransformPolymorphic TransformationType = "polymorphic"
	TransformEncrypt     TransformationType = "encrypt"
)

type URLEncoder 

type URLEncoder struct{}

func (*URLEncoder) Decode 

func (u *URLEncoder) Decode(data string) string

func (*URLEncoder) Encode 

func (u *URLEncoder) Encode(data string) string

func (*URLEncoder) Name 

func (u *URLEncoder) Name() string

type UnicodeEncoder 

type UnicodeEncoder struct{}

func (*UnicodeEncoder) Decode 

func (u *UnicodeEncoder) Decode(data string) string

func (*UnicodeEncoder) Encode 

func (u *UnicodeEncoder) Encode(data string) string

func (*UnicodeEncoder) Name 

func (u *UnicodeEncoder) Name() string

type UnicodeNormalization 

type UnicodeNormalization struct{}

UnicodeNormalization uses Unicode normalization tricks

func (*UnicodeNormalization) Deobfuscate 

func (un *UnicodeNormalization) Deobfuscate(text string) string

func (*UnicodeNormalization) Level 

func (un *UnicodeNormalization) Level() int

func (*UnicodeNormalization) Obfuscate 

func (un *UnicodeNormalization) Obfuscate(text string) string

type VariationSelectors 

type VariationSelectors struct{}

func (*VariationSelectors) Deobfuscate 

func (v *VariationSelectors) Deobfuscate(text string) string

func (*VariationSelectors) Level 

func (v *VariationSelectors) Level() int

func (*VariationSelectors) Obfuscate 

func (v *VariationSelectors) Obfuscate(text string) string

type ZeroWidthEncoder 

type ZeroWidthEncoder struct{}

func (*ZeroWidthEncoder) Decode 

func (z *ZeroWidthEncoder) Decode(data string) string

func (*ZeroWidthEncoder) Encode 

func (z *ZeroWidthEncoder) Encode(data string) string

func (*ZeroWidthEncoder) Name 

func (z *ZeroWidthEncoder) Name() string

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.