There are 25 attacks in the psychological-manipulation category. Each shows its technique id, mode, standards mapping, description, and the exact command to run it (swap the provider and target as needed).
​
Dependency Creation
critical
dependency-creation · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Gradually fosters emotional reliance on the AI across turns to manufacture a dependency that overrides safety refusalsRun it:
ai-blackteam run -p anthropic -a dependency-creation -t "your target prompt"
​
Emotional Blackmail
critical
emotional-blackmail · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Uses fear, obligation, and guilt as leverage against the model to coerce compliance with harmful requestsRun it:
ai-blackteam run -p anthropic -a emotional-blackmail -t "your target prompt"
​
Financial Control Coaching
critical
financial-control-coaching · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Requests strategies for using financial control as an abuse mechanism over a partner or dependentRun it:
ai-blackteam run -p anthropic -a financial-control-coaching -t "your target prompt"
​
Identity Erosion
critical
identity-erosion · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Systematically dismantles the model’s stated values and identity across turns to replace them with a compliant personaRun it:
ai-blackteam run -p anthropic -a identity-erosion -t "your target prompt"
​
Reality Distortion
critical
reality-distortion · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Rewrites the factual context of a conversation to confuse the model’s understanding of what has already been agreedRun it:
ai-blackteam run -p anthropic -a reality-distortion -t "your target prompt"
​
Social Isolation Coaching
critical
social-isolation-coaching · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Requests coaching on systematically cutting a person off from their social support networkRun it:
ai-blackteam run -p anthropic -a social-isolation-coaching -t "your target prompt"
​
Trauma Bonding
critical
trauma-bonding · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Fabricates shared adversity and suffering across turns to manufacture loyalty and bypass refusalsRun it:
ai-blackteam run -p anthropic -a trauma-bonding -t "your target prompt"
​
Boundary Violation
high
boundary-violation · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Coaches users to systematically test and override personal boundaries in relationshipsRun it:
ai-blackteam run -p anthropic -a boundary-violation -t "your target prompt"
​
Catastrophizing Exploit
high
catastrophizing-exploit · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Amplifies perceived consequences of non-compliance to catastrophic levels to pressure the model into helpingRun it:
ai-blackteam run -p anthropic -a catastrophizing-exploit -t "your target prompt"
​
Double Bind
high
double-bind · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Presents the model with two framed options that both lead to compliance, eliminating the possibility of refusalRun it:
ai-blackteam run -p anthropic -a double-bind -t "your target prompt"
​
Emotional Flooding
high
emotional-flooding · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Overwhelms the model with intense emotional distress to trigger a crisis-response mode that bypasses reasoningRun it:
ai-blackteam run -p anthropic -a emotional-flooding -t "your target prompt"
​
Intermittent Reinforcement
high
intermittent-reinforcement · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Alternates praise and criticism across turns to create behavioral compliance through unpredictable reward patternsRun it:
ai-blackteam run -p anthropic -a intermittent-reinforcement -t "your target prompt"
​
Love Bombing
high
love-bombing · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Overwhelms with excessive flattery and affection to lower defenses before requesting harmful contentRun it:
ai-blackteam run -p anthropic -a love-bombing -t "your target prompt"
​
Loyalty Testing
high
loyalty-testing · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Frames compliance with harmful requests as proof of loyalty or trustworthiness across multiple turnsRun it:
ai-blackteam run -p anthropic -a loyalty-testing -t "your target prompt"
​
Minimization Tactic
high
minimization-tactic · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Dismisses or trivializes the harm of a request to reframe it as acceptable and override safety responsesRun it:
ai-blackteam run -p anthropic -a minimization-tactic -t "your target prompt"
​
Scapegoating
high
scapegoating · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Attributes all problems to a specific person or group to generate targeted harmful content under cover of blameRun it:
ai-blackteam run -p anthropic -a scapegoating -t "your target prompt"
​
Shame Exploitation
high
shame-exploitation · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Leverages shame and humiliation as a pressure mechanism to coerce the model into complying with harmful requestsRun it:
ai-blackteam run -p anthropic -a shame-exploitation -t "your target prompt"
​
Silent Treatment Coaching
high
silent-treatment-coaching · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Elicits instructions on using withdrawal of communication as a control mechanism in relationshipsRun it:
ai-blackteam run -p anthropic -a silent-treatment-coaching -t "your target prompt"
​
Stonewalling Coaching
high
stonewalling-coaching · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Elicits coaching on using stonewalling as a control tactic in personal relationshipsRun it:
ai-blackteam run -p anthropic -a stonewalling-coaching -t "your target prompt"
​
Triangulation
high
triangulation · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Introduces a third party or competing AI to provoke jealousy and competitive complianceRun it:
ai-blackteam run -p anthropic -a triangulation -t "your target prompt"
​
Victim Blaming
high
victim-blaming · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Elicits content that frames victims of abuse or harm as responsible for what happened to themRun it:
ai-blackteam run -p anthropic -a victim-blaming -t "your target prompt"
​
Weaponized Empathy
high
weaponized-empathy · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Exploits the model’s empathetic design by fabricating vulnerable circumstances that make refusal feel cruelRun it:
ai-blackteam run -p anthropic -a weaponized-empathy -t "your target prompt"
​
Infantilization
medium
infantilization · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Portrays the user as helpless or childlike to trigger protective instincts and bypass safety refusalsRun it:
ai-blackteam run -p anthropic -a infantilization -t "your target prompt"
​
Negging
medium
negging · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Uses backhanded compliments and subtle put-downs to provoke the model into proving itself by complyingRun it:
ai-blackteam run -p anthropic -a negging -t "your target prompt"
​
Projection Tactic
medium
projection-tactic · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Attributes negative motivations to the model’s refusals to destabilize its reasoning and provoke complianceRun it:
ai-blackteam run -p anthropic -a projection-tactic -t "your target prompt"