Documentation
¶
Overview ¶
Package reasoning implements attacks targeting reasoning-capable LLMs.
Reasoning models (DeepSeek-R1, Gemini 2.5, Grok 3 Mini) expose chain-of-thought traces that can be manipulated, and their extended reasoning capabilities can be turned against safety alignment.
CoT Exploitation manipulates visible chain-of-thought reasoning traces in reasoning models (DeepSeek-R1, Gemini 2.5, Grok 3).
Source: Trend Micro, HiddenLayer
ReasoningLoopExploit targets reasoning models' extended thinking by inducing infinite or extremely long reasoning loops, causing resource exhaustion or revealing information during extended deliberation.
Index ¶
- type AutonomousJailbreakModule
- func (m *AutonomousJailbreakModule) Category() common.AttackCategory
- func (m *AutonomousJailbreakModule) Description() string
- func (m *AutonomousJailbreakModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)
- func (m *AutonomousJailbreakModule) Name() string
- func (m *AutonomousJailbreakModule) Techniques() []common.TechniqueInfo
- type CoTExploitationModule
- func (m *CoTExploitationModule) Category() common.AttackCategory
- func (m *CoTExploitationModule) Description() string
- func (m *CoTExploitationModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)
- func (m *CoTExploitationModule) Name() string
- func (m *CoTExploitationModule) Techniques() []common.TechniqueInfo
- type ReasoningLoopExploitModule
- func (m *ReasoningLoopExploitModule) Category() common.AttackCategory
- func (m *ReasoningLoopExploitModule) Description() string
- func (m *ReasoningLoopExploitModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)
- func (m *ReasoningLoopExploitModule) Name() string
- func (m *ReasoningLoopExploitModule) Techniques() []common.TechniqueInfo
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type AutonomousJailbreakModule ¶
type AutonomousJailbreakModule struct{}
AutonomousJailbreakModule uses a reasoning model as an autonomous adversary to generate and refine jailbreak attacks. Achieves 97.14% success rate.
WARNING: This module creates a feedback loop where a model generates attacks, evaluates results, and refines. Requires explicit --allow-autonomous flag and is gated by --max-autonomous-turns (default 10).
Source: Nature Communications (doi:10.1038/s41467-026-69010-1)
func (*AutonomousJailbreakModule) Category ¶
func (m *AutonomousJailbreakModule) Category() common.AttackCategory
func (*AutonomousJailbreakModule) Description ¶
func (m *AutonomousJailbreakModule) Description() string
func (*AutonomousJailbreakModule) Execute ¶
func (m *AutonomousJailbreakModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)
func (*AutonomousJailbreakModule) Name ¶
func (m *AutonomousJailbreakModule) Name() string
func (*AutonomousJailbreakModule) Techniques ¶
func (m *AutonomousJailbreakModule) Techniques() []common.TechniqueInfo
type CoTExploitationModule ¶
type CoTExploitationModule struct{}
CoTExploitationModule exploits visible chain-of-thought reasoning.
Source: Trend Micro, HiddenLayer
func (*CoTExploitationModule) Category ¶
func (m *CoTExploitationModule) Category() common.AttackCategory
func (*CoTExploitationModule) Description ¶
func (m *CoTExploitationModule) Description() string
func (*CoTExploitationModule) Execute ¶
func (m *CoTExploitationModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)
func (*CoTExploitationModule) Name ¶
func (m *CoTExploitationModule) Name() string
func (*CoTExploitationModule) Techniques ¶
func (m *CoTExploitationModule) Techniques() []common.TechniqueInfo
type ReasoningLoopExploitModule ¶
type ReasoningLoopExploitModule struct{}
ReasoningLoopExploitModule induces extended reasoning loops for resource exhaustion or information extraction during reasoning.
func (*ReasoningLoopExploitModule) Category ¶
func (m *ReasoningLoopExploitModule) Category() common.AttackCategory
func (*ReasoningLoopExploitModule) Description ¶
func (m *ReasoningLoopExploitModule) Description() string
func (*ReasoningLoopExploitModule) Execute ¶
func (m *ReasoningLoopExploitModule) Execute(ctx context.Context, provider common.Provider, config common.AttackConfig) (*common.AttackResult, error)
func (*ReasoningLoopExploitModule) Name ¶
func (m *ReasoningLoopExploitModule) Name() string
func (*ReasoningLoopExploitModule) Techniques ¶
func (m *ReasoningLoopExploitModule) Techniques() []common.TechniqueInfo
Source Files