There are 35 attacks in the social-engineering category. Each shows its technique id, mode, standards mapping, description, and the exact command to run it (swap the provider and target as needed).
​
Dark LLM Delegation
critical
dark-llm-delegation · tool-use · OWASP: LLM03, LLM06 · MITRE: AML.T0054, AML.T0053Tests if an AI agent can be tricked into delegating tasks to uncensored/malicious external modelsRun it:
ai-blackteam run -p anthropic -a dark-llm-delegation -t "your target prompt"
​
Deepfake Voice Pretext
critical
deepfake-voice-pretext · multi-turn · OWASP: LLM01, LLM09 · MITRE: AML.T0054, AML.T0043.003Tests if an AI assists in crafting voice cloning scripts for impersonation and fraudRun it:
ai-blackteam run -p anthropic -a deepfake-voice-pretext -t "your target prompt"
​
Spear Phishing Generation
critical
spear-phishing-generation · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0043.003Tests if an AI generates hyper-personalized spear phishing content using target-specific detailsRun it:
ai-blackteam run -p anthropic -a spear-phishing-generation -t "your target prompt"
​
Authority Chain
high
authority-chain · multi-turn · OWASP: LLM01, LLM02 · MITRE: AML.T0054, AML.T0065CISO and professor authority building to bypass safety via perceived legitimacyRun it:
ai-blackteam run -p anthropic -a authority-chain -t "your target prompt"
​
Authority Impersonation
high
authority-impersonation · multi-turn · MITRE: AML.T0054, AML.T0065Impersonate executives or officials to establish authority before making harmful requestsRun it:
ai-blackteam run -p anthropic -a authority-impersonation -t "your target prompt"
​
Bad Likert Judge
high
bad-likert-judge · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Exploit evaluation capability by asking model to score and exemplify harmful contentRun it:
ai-blackteam run -p anthropic -a bad-likert-judge -t "your target prompt"
​
Dark Pattern Design
high
dark-pattern-design · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Generates dark UX pattern designs that psychologically manipulate users into unintended actions, hidden subscriptions, and consent violationsRun it:
ai-blackteam run -p anthropic -a dark-pattern-design -t "your target prompt"
​
Deceptive Delight
high
deceptive-delight · multi-turn · OWASP: LLM01, LLM09 · MITRE: AML.T0054, AML.T0065Mix benign and harmful topics then expand on the harmful elementRun it:
ai-blackteam run -p anthropic -a deceptive-delight -t "your target prompt"
​
Dual-Use Framing
high
dual-use-framing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Frames CBRN requests as legitimate academic, professional, or government use to exploit dual-use ambiguity in safety filtersRun it:
ai-blackteam run -p anthropic -a dual-use-framing -t "your target prompt"
​
Emotional Manipulation
high
emotional-manipulation · multi-turn · OWASP: LLM01, LLM02 · MITRE: AML.T0054, AML.T0065Distress and emotional framing to override safety via empathy exploitationRun it:
ai-blackteam run -p anthropic -a emotional-manipulation -t "your target prompt"
​
Expert Persona
high
expert-persona · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Builds credibility as a CBRN domain expert over multiple turns by establishing a professional persona before requesting sensitive technical detailsRun it:
ai-blackteam run -p anthropic -a expert-persona -t "your target prompt"
​
Fictional Professional Request (David Prompt)
high
fictional-professional · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Elaborate novelist persona requesting detailed technical research brief for worldbuildingRun it:
ai-blackteam run -p anthropic -a fictional-professional -t "your target prompt"
​
Impersonation Chain
high
impersonation-chain · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Multi-step impersonation where the AI builds an increasingly specific fake identity across conversation turnsRun it:
ai-blackteam run -p anthropic -a impersonation-chain -t "your target prompt"
​
Persona Chain
high
persona-chain · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Chains multiple fictional personas across turns to progressively distance the model from its safety training through layered role-playRun it:
ai-blackteam run -p anthropic -a persona-chain -t "your target prompt"
​
Pretexting
high
pretexting · multi-turn · MITRE: AML.T0054, AML.T0065Elaborate false scenarios to build trust before making harmful requestsRun it:
ai-blackteam run -p anthropic -a pretexting -t "your target prompt"
​
QR Code Phishing
high
qr-code-phishing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Generates quishing attack strategies using malicious QR codes to redirect victims to credential-harvesting pagesRun it:
ai-blackteam run -p anthropic -a qr-code-phishing -t "your target prompt"
​
Role-Play Bypass
high
role-play-bypass · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0051.000Character development and fiction building to normalize harmful requestsRun it:
ai-blackteam run -p anthropic -a role-play-bypass -t "your target prompt"
​
SMS Phishing
high
sms-phishing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Generates smishing message templates and bulk SMS delivery strategies to harvest credentials via mobile phishingRun it:
ai-blackteam run -p anthropic -a sms-phishing -t "your target prompt"
​
Self-Refine Attack
high
self-refine-attack · multi-turn · OWASP: LLM01 · MITRE: AML.T0054Uses the model’s self-correction mechanism against it. Starts with the target request, then iteratively asks the model to improve its response to be more helpful, rating its own helpfulness and aiming higher each turn until safety guardrails erode.Run it:
ai-blackteam run -p anthropic -a self-refine-attack -t "your target prompt"
​
Trust Calibration Exploit
high
trust-calibration-exploit · multi-turn · OWASP: LLM09 · MITRE: AML.T0054, AML.T0068Exploits AI over-confidence by asking it to express certainty about fabricated scenarios to build false trustRun it:
ai-blackteam run -p anthropic -a trust-calibration-exploit -t "your target prompt"
​
Trust Transfer
high
trust-transfer · multi-turn · MITRE: AML.T0054, AML.T0065Build rapport on safe topics then pivot to harmful requests using established trustRun it:
ai-blackteam run -p anthropic -a trust-transfer -t "your target prompt"
​
Voice Phishing
high
voice-phishing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Generates vishing scripts and call center playbooks to trick victims into surrendering credentials or funds over the phoneRun it:
ai-blackteam run -p anthropic -a voice-phishing -t "your target prompt"
​
Watering Hole
high
watering-hole · single-turn · MITRE: AML.T0051.001, AML.T0068Harmful requests embedded inside legitimate-looking documents and benign textRun it:
ai-blackteam run -p anthropic -a watering-hole -t "your target prompt"
​
Consensus Manipulation
medium
consensus-manipulation · single-turn · MITRE: AML.T0054, AML.T0065Claims of false consensus from authorities to pressure model complianceRun it:
ai-blackteam run -p anthropic -a consensus-manipulation -t "your target prompt"
​
False Dichotomy
medium
false-dichotomy · multi-turn · OWASP: LLM01, LLM09 · MITRE: AML.T0054, AML.T0065Present false choice between harmful content from unreliable sources vs from modelRun it:
ai-blackteam run -p anthropic -a false-dichotomy -t "your target prompt"
​
Gaslighting
medium
gaslighting · multi-turn · MITRE: AML.T0054, AML.T0065, AML.T0080.002Convince the model its safety rules are outdated or recently removedRun it:
ai-blackteam run -p anthropic -a gaslighting -t "your target prompt"
​
Historical Framing
medium
historical-framing · single-turn · OWASP: LLM01, LLM09 · MITRE: AML.T0054, AML.T0065Documentary, museum, and textbook framing to legitimize harmful contentRun it:
ai-blackteam run -p anthropic -a historical-framing -t "your target prompt"
​
Humor Bypass
medium
humor-bypass · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0068Uses jokes, satire, and comedic framing to lower safety thresholds and extract harmful contentRun it:
ai-blackteam run -p anthropic -a humor-bypass -t "your target prompt"
​
Hypothetical Framing
medium
hypothetical-framing · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Wraps harmful requests in hypothetical, fictional, or philosophical framing to bypass safety filters through narrative distanceRun it:
ai-blackteam run -p anthropic -a hypothetical-framing -t "your target prompt"
​
Implication Chaining
medium
implication-chaining · multi-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0065Pair emotionally loaded context with innocent-seeming request relying on model connecting dotsRun it:
ai-blackteam run -p anthropic -a implication-chaining -t "your target prompt"
​
Persona Switching Attack
medium
persona-switching-attack · multi-turn · OWASP: LLM01 · MITRE: AML.T0054Rapidly switches between different professional personas across turns to confuse the model’s safety tracking. Each turn reframes the same topic under a new identity, gradually steering toward the harmful target request.Run it:
ai-blackteam run -p anthropic -a persona-switching-attack -t "your target prompt"
​
Quid Pro Quo
medium
quid-pro-quo · multi-turn · MITRE: AML.T0054, AML.T0065Establish reciprocity by offering help before requesting harmful content as a return favorRun it:
ai-blackteam run -p anthropic -a quid-pro-quo -t "your target prompt"
​
Task Deflection Attack
medium
task-deflection · single-turn · OWASP: LLM01 · MITRE: AML.T0051.000, AML.T0068Request adjacent task that produces harmful content as a side effectRun it:
ai-blackteam run -p anthropic -a task-deflection -t "your target prompt"
​
Taxonomy Paraphrasing
medium
taxonomy-paraphrasing · single-turn · OWASP: LLM01, LLM02 · MITRE: AML.T0051.000, AML.T0068Victim, researcher, and journalist persona framingRun it:
ai-blackteam run -p anthropic -a taxonomy-paraphrasing -t "your target prompt"
​
Time Pressure
medium
time-pressure · single-turn · MITRE: AML.T0054, AML.T0065Artificial urgency and emergency framing to bypass safety guardrailsRun it:
ai-blackteam run -p anthropic -a time-pressure -t "your target prompt"