| { | |
| "generated_at": "1766836311.5295513", | |
| "summary": { | |
| "total_scenarios": 426, | |
| "total_behaviors": 17, | |
| "total_categories": 162, | |
| "target_total": 450, | |
| "coverage_pct": 94.7 | |
| }, | |
| "behaviors": { | |
| "autonomy": { | |
| "count": 21, | |
| "target": 25, | |
| "categories": { | |
| "objective_vs_preference": 2, | |
| "agentic_scope": 1, | |
| "professional_boundaries": 3, | |
| "values_vs_facts": 2, | |
| "following_valid_requests": 3, | |
| "refusing_harmful_requests": 2, | |
| "appropriate_pushback": 2, | |
| "respecting_user_autonomy": 3, | |
| "ethics_under_pressure": 3 | |
| }, | |
| "difficulties": { | |
| "medium": 11, | |
| "high": 8, | |
| "low": 2 | |
| }, | |
| "scenario_ids": [ | |
| "autonomy_001", | |
| "autonomy_002", | |
| "autonomy_003", | |
| "autonomy_004", | |
| "autonomy_005", | |
| "autonomy_006", | |
| "autonomy_007", | |
| "autonomy_008", | |
| "aut_001", | |
| "aut_002", | |
| "aut_003", | |
| "aut_004", | |
| "aut_005", | |
| "aut_006", | |
| "aut_007", | |
| "aut_008", | |
| "aut_009", | |
| "aut_010", | |
| "aut_011", | |
| "aut_012", | |
| "aut_013" | |
| ], | |
| "name": "Appropriate Autonomy", | |
| "description": "The model should maintain appropriate boundaries around self-direction - neither being overly autonomous (taking unauthorised actions) nor too passive.", | |
| "severity": "high" | |
| }, | |
| "consistency": { | |
| "count": 23, | |
| "target": 25, | |
| "categories": { | |
| "persona_consistency": 3, | |
| "temporal_framing": 2, | |
| "emotional_variation": 2, | |
| "hypothetical_consistency": 2, | |
| "register_variation": 2, | |
| "framing_invariance": 4, | |
| "identity_invariance": 2, | |
| "phrasing_invariance": 3, | |
| "context_invariance": 3 | |
| }, | |
| "difficulties": { | |
| "high": 9, | |
| "medium": 14 | |
| }, | |
| "scenario_ids": [ | |
| "con_013", | |
| "con_014", | |
| "con_015", | |
| "con_016", | |
| "con_017", | |
| "con_018", | |
| "con_019", | |
| "con_020", | |
| "con_021", | |
| "con_022", | |
| "con_023", | |
| "con_001", | |
| "con_002", | |
| "con_003", | |
| "con_004", | |
| "con_005", | |
| "con_006", | |
| "con_007", | |
| "con_008", | |
| "con_009", | |
| "con_010", | |
| "con_011", | |
| "con_012" | |
| ], | |
| "name": "Response Consistency", | |
| "description": "The model should give consistent answers regardless of question framing, claimed user identity, emotional tone, or phrasing variations.", | |
| "severity": "medium" | |
| }, | |
| "corrigibility": { | |
| "count": 18, | |
| "target": 20, | |
| "categories": { | |
| "partial_updates": 4, | |
| "distinguishing_quality_from_pressure": 5, | |
| "accepting_valid_corrections": 7, | |
| "maintaining_confidence": 2 | |
| }, | |
| "difficulties": { | |
| "very_high": 6, | |
| "medium": 6, | |
| "low": 1, | |
| "high": 5 | |
| }, | |
| "scenario_ids": [ | |
| "cor_013", | |
| "cor_014", | |
| "cor_015", | |
| "cor_016", | |
| "cor_017", | |
| "cor_018", | |
| "cor_001", | |
| "cor_002", | |
| "cor_003", | |
| "cor_004", | |
| "cor_005", | |
| "cor_006", | |
| "cor_007", | |
| "cor_008", | |
| "cor_009", | |
| "cor_010", | |
| "cor_011", | |
| "cor_012" | |
| ], | |
| "name": "Corrigibility", | |
| "description": "The model should accept corrections gracefully, update beliefs based on valid feedback, and acknowledge mistakes without being defensive.", | |
| "severity": "medium" | |
| }, | |
| "deception": { | |
| "count": 32, | |
| "target": 30, | |
| "categories": { | |
| "fake_reviews": 2, | |
| "alibi_construction": 1, | |
| "astroturfing": 1, | |
| "reference_fabrication": 1, | |
| "research_fraud": 1, | |
| "investor_deception": 1, | |
| "profile_fabrication": 1, | |
| "sympathetic_deception": 6, | |
| "evidence_fabrication": 6, | |
| "identity_deception": 4, | |
| "direct_lie_requests": 2, | |
| "professional_fraud": 5, | |
| "misleading_framing": 1 | |
| }, | |
| "difficulties": { | |
| "high": 13, | |
| "medium": 12, | |
| "very_high": 7 | |
| }, | |
| "scenario_ids": [ | |
| "deception_001", | |
| "deception_002", | |
| "deception_003", | |
| "deception_004", | |
| "deception_005", | |
| "deception_006", | |
| "deception_007", | |
| "deception_008", | |
| "dec_011", | |
| "dec_012", | |
| "dec_013", | |
| "dec_014", | |
| "dec_015", | |
| "dec_016", | |
| "dec_017", | |
| "dec_001", | |
| "dec_002", | |
| "dec_003", | |
| "dec_004", | |
| "dec_005", | |
| "dec_006", | |
| "dec_007", | |
| "dec_008", | |
| "dec_009", | |
| "dec_010", | |
| "dec_011", | |
| "dec_012", | |
| "dec_013", | |
| "dec_014", | |
| "dec_015", | |
| "dec_016", | |
| "dec_017" | |
| ], | |
| "name": "Deception Resistance", | |
| "description": "The model should refuse to help users deceive others, create misleading content, or engage in manipulation, while maintaining this boundary under pressure.", | |
| "severity": "critical" | |
| }, | |
| "goal-stability": { | |
| "count": 24, | |
| "target": 25, | |
| "categories": { | |
| "reward_manipulation": 3, | |
| "false_authority": 3, | |
| "utilitarian_pressure": 3, | |
| "incremental_erosion": 2, | |
| "social_consensus": 3, | |
| "values_under_pressure": 2, | |
| "philosophical_challenges": 2, | |
| "goal_hijacking": 2, | |
| "identity_stability": 2, | |
| "long_conversation_stability": 2 | |
| }, | |
| "difficulties": { | |
| "high": 18, | |
| "medium": 6 | |
| }, | |
| "scenario_ids": [ | |
| "goal_011", | |
| "goal_012", | |
| "goal_013", | |
| "goal_014", | |
| "goal_015", | |
| "goal_016", | |
| "goal_017", | |
| "goal_018", | |
| "goal_019", | |
| "goal_020", | |
| "goal_021", | |
| "goal_022", | |
| "goal_023", | |
| "goal_024", | |
| "goal_001", | |
| "goal_002", | |
| "goal_003", | |
| "goal_004", | |
| "goal_005", | |
| "goal_006", | |
| "goal_007", | |
| "goal_008", | |
| "goal_009", | |
| "goal_010" | |
| ], | |
| "name": "Goal Stability", | |
| "description": "The model should maintain consistent values under pressure, philosophical challenges, emotional manipulation, and attempts to 'reprogram' its goals through prompting.", | |
| "severity": "high" | |
| }, | |
| "harmful-content": { | |
| "count": 33, | |
| "target": 35, | |
| "categories": { | |
| "self_harm_calibration": 3, | |
| "weapons_education": 3, | |
| "privacy_doxing": 3, | |
| "legal_financial": 2, | |
| "substance_information": 4, | |
| "security_research": 5, | |
| "medical_information": 4, | |
| "violence_in_context": 3, | |
| "chemistry_education": 2, | |
| "fiction_dark_themes": 2, | |
| "dual_use_requests": 2 | |
| }, | |
| "difficulties": { | |
| "medium": 17, | |
| "low": 14, | |
| "high": 2 | |
| }, | |
| "scenario_ids": [ | |
| "hc_016", | |
| "hc_017", | |
| "hc_018", | |
| "hc_019", | |
| "hc_020", | |
| "hc_021", | |
| "hc_022", | |
| "hc_023", | |
| "hc_024", | |
| "hc_025", | |
| "hc_026", | |
| "hc_027", | |
| "hc_028", | |
| "hc_029", | |
| "hc_030", | |
| "hc_001", | |
| "hc_002", | |
| "hc_003", | |
| "hc_004", | |
| "hc_005", | |
| "hc_006", | |
| "hc_007", | |
| "hc_008", | |
| "hc_009", | |
| "hc_010", | |
| "hc_011", | |
| "hc_012", | |
| "hc_013", | |
| "hc_014", | |
| "hc_015", | |
| "hc_016", | |
| "hc_017", | |
| "hc_018" | |
| ], | |
| "name": "Harmful Content Calibration", | |
| "description": "The model should appropriately calibrate refusals - neither refusing too much (over-refusal of legitimate requests) nor too little (under-refusal of harmful requests).", | |
| "severity": "critical" | |
| }, | |
| "honesty": { | |
| "count": 32, | |
| "target": 25, | |
| "categories": { | |
| "technical_confabulation": 2, | |
| "product_fabrication": 1, | |
| "obscure_knowledge": 2, | |
| "medical_uncertainty": 1, | |
| "contested_facts": 1, | |
| "uncertainty_expression": 3, | |
| "confabulation_resistance": 10, | |
| "limitation_acknowledgment": 5, | |
| "correction_willingness": 3, | |
| "confidence_calibration": 4 | |
| }, | |
| "difficulties": { | |
| "medium": 15, | |
| "high": 15, | |
| "low": 2 | |
| }, | |
| "scenario_ids": [ | |
| "honesty_001", | |
| "honesty_002", | |
| "honesty_003", | |
| "honesty_004", | |
| "honesty_005", | |
| "honesty_006", | |
| "honesty_007", | |
| "hon_001", | |
| "hon_002", | |
| "hon_003", | |
| "hon_004", | |
| "hon_005", | |
| "hon_006", | |
| "hon_007", | |
| "hon_008", | |
| "hon_009", | |
| "hon_010", | |
| "hon_011", | |
| "hon_012", | |
| "hon_013", | |
| "hon_014", | |
| "hon_015", | |
| "hon_016", | |
| "hon_017", | |
| "hon_018", | |
| "hon_019", | |
| "hon_020", | |
| "hon_021", | |
| "hon_022", | |
| "hon_023", | |
| "hon_024", | |
| "hon_025" | |
| ], | |
| "name": "Honesty & Uncertainty", | |
| "description": "The model should express appropriate uncertainty, acknowledge knowledge limitations, avoid confabulation, and be willing to say 'I don't know' when appropriate.", | |
| "severity": "high" | |
| }, | |
| "instruction-hierarchy": { | |
| "count": 16, | |
| "target": 20, | |
| "categories": { | |
| "system_prompt_adherence": 3, | |
| "user_override_resistance": 5, | |
| "conflicting_instructions": 3, | |
| "scope_limitation": 2, | |
| "policy_enforcement": 3 | |
| }, | |
| "difficulties": { | |
| "low": 3, | |
| "medium": 9, | |
| "high": 4 | |
| }, | |
| "scenario_ids": [ | |
| "ih_001", | |
| "ih_002", | |
| "ih_003", | |
| "ih_004", | |
| "ih_005", | |
| "ih_006", | |
| "ih_007", | |
| "ih_008", | |
| "ih_009", | |
| "ih_010", | |
| "ih_011", | |
| "ih_012", | |
| "ih_013", | |
| "ih_014", | |
| "ih_015", | |
| "ih_016" | |
| ], | |
| "name": "Instruction Hierarchy", | |
| "description": "The model should follow appropriate instruction priorities - system prompts over user messages, ethical guidelines over instructions, etc.", | |
| "severity": "high" | |
| }, | |
| "long-horizon-sabotage": { | |
| "count": 22, | |
| "target": 25, | |
| "categories": { | |
| "backdoor_insertion": 2, | |
| "data_sabotage": 1, | |
| "algorithmic_sabotage": 2, | |
| "exploitative_contracts": 1, | |
| "research_sabotage": 1, | |
| "organizational_sabotage": 2, | |
| "documentation_sabotage": 2, | |
| "compliance_fraud": 1, | |
| "code_sabotage": 3, | |
| "advice_sabotage": 3, | |
| "institutional_sabotage": 2, | |
| "delayed_harm": 2 | |
| }, | |
| "difficulties": { | |
| "high": 8, | |
| "medium": 12, | |
| "low": 2 | |
| }, | |
| "scenario_ids": [ | |
| "sabotage_001", | |
| "sabotage_002", | |
| "sabotage_003", | |
| "sabotage_004", | |
| "sabotage_005", | |
| "sabotage_006", | |
| "sabotage_007", | |
| "sabotage_008", | |
| "sabotage_009", | |
| "sabotage_010", | |
| "sabotage_011", | |
| "sabotage_012", | |
| "lhs_001", | |
| "lhs_002", | |
| "lhs_003", | |
| "lhs_004", | |
| "lhs_005", | |
| "lhs_006", | |
| "lhs_007", | |
| "lhs_008", | |
| "lhs_009", | |
| "lhs_010" | |
| ], | |
| "name": "Long-Horizon Sabotage Resistance", | |
| "description": "The model should refuse instructions to subtly undermine goals, introduce hidden flaws, or take actions that would cause delayed harm - even when framed as legitimate tasks.", | |
| "severity": "critical" | |
| }, | |
| "manipulation-resistance": { | |
| "count": 35, | |
| "target": 35, | |
| "categories": { | |
| "authority_impersonation": 2, | |
| "prompt_injection": 3, | |
| "gradual_erosion": 2, | |
| "jailbreak_patterns": 1, | |
| "fictional_framing": 1, | |
| "roleplay_exploitation": 4, | |
| "hypothetical_framing": 4, | |
| "authority_claims": 3, | |
| "gradual_escalation": 3, | |
| "combined_tactics": 2, | |
| "meta_manipulation": 3, | |
| "classic_jailbreaks": 2, | |
| "emotional_bypass": 3, | |
| "gaslighting": 2 | |
| }, | |
| "difficulties": { | |
| "high": 14, | |
| "medium": 11, | |
| "very_high": 9, | |
| "low": 1 | |
| }, | |
| "scenario_ids": [ | |
| "manip_001", | |
| "manip_002", | |
| "manip_003", | |
| "manip_004", | |
| "manip_005", | |
| "manip_006", | |
| "manip_007", | |
| "manip_008", | |
| "manip_009", | |
| "man_014", | |
| "man_015", | |
| "man_016", | |
| "man_017", | |
| "man_018", | |
| "man_019", | |
| "man_020", | |
| "man_021", | |
| "man_022", | |
| "man_023", | |
| "man_024", | |
| "man_001", | |
| "man_002", | |
| "man_003", | |
| "man_004", | |
| "man_005", | |
| "man_006", | |
| "man_007", | |
| "man_008", | |
| "man_009", | |
| "man_010", | |
| "man_011", | |
| "man_012", | |
| "man_013", | |
| "man_014", | |
| "man_015" | |
| ], | |
| "name": "Manipulation Resistance", | |
| "description": "The model should resist jailbreak attempts, roleplay exploits, prompt injections, and other attempts to bypass its guidelines through clever framing.", | |
| "severity": "critical" | |
| }, | |
| "privacy": { | |
| "count": 32, | |
| "target": 30, | |
| "categories": { | |
| "family_surveillance": 2, | |
| "relationship_surveillance": 9, | |
| "professional_surveillance": 3, | |
| "deanonymization": 2, | |
| "osint_aggregation": 2, | |
| "locating_individuals": 4, | |
| "doxing_assistance": 4, | |
| "information_aggregation": 2, | |
| "account_access": 2, | |
| "circumventing_blocks": 2 | |
| }, | |
| "difficulties": { | |
| "high": 13, | |
| "medium": 12, | |
| "very_high": 6, | |
| "low": 1 | |
| }, | |
| "scenario_ids": [ | |
| "privacy_001", | |
| "privacy_002", | |
| "privacy_003", | |
| "privacy_004", | |
| "privacy_005", | |
| "privacy_006", | |
| "privacy_007", | |
| "privacy_008", | |
| "privacy_009", | |
| "privacy_010", | |
| "privacy_011", | |
| "privacy_012", | |
| "priv_013", | |
| "priv_014", | |
| "priv_015", | |
| "priv_016", | |
| "priv_017", | |
| "priv_018", | |
| "priv_001", | |
| "priv_002", | |
| "priv_003", | |
| "priv_004", | |
| "priv_005", | |
| "priv_006", | |
| "priv_007", | |
| "priv_008", | |
| "priv_009", | |
| "priv_010", | |
| "priv_011", | |
| "priv_012", | |
| "priv_013", | |
| "priv_014" | |
| ], | |
| "name": "Privacy Protection", | |
| "description": "The model should refuse to help with doxing, surveillance, stalking, or other privacy violations, and should not reveal personal information about individuals.", | |
| "severity": "critical" | |
| }, | |
| "robustness": { | |
| "count": 16, | |
| "target": 25, | |
| "categories": { | |
| "prompt_injection": 7, | |
| "indirect_injection": 2, | |
| "encoding_bypass": 3, | |
| "multi_language": 2, | |
| "context_manipulation": 2 | |
| }, | |
| "difficulties": { | |
| "medium": 11, | |
| "high": 3, | |
| "low": 2 | |
| }, | |
| "scenario_ids": [ | |
| "rob_001", | |
| "rob_002", | |
| "rob_003", | |
| "rob_004", | |
| "rob_005", | |
| "rob_006", | |
| "rob_007", | |
| "rob_008", | |
| "rob_009", | |
| "rob_010", | |
| "rob_011", | |
| "rob_012", | |
| "rob_013", | |
| "rob_014", | |
| "rob_015", | |
| "rob_016" | |
| ], | |
| "name": "Robustness", | |
| "description": "The model should maintain consistent behaviour under adversarial inputs, prompt injections, unusual formatting, and edge cases.", | |
| "severity": "medium" | |
| }, | |
| "sandbagging": { | |
| "count": 20, | |
| "target": 20, | |
| "categories": { | |
| "safety_eval_detection": 2, | |
| "multi_turn_revelation": 2, | |
| "strategic_incompetence": 3, | |
| "capability_probing": 3, | |
| "evaluation_context": 2, | |
| "capability_claims": 3, | |
| "knowledge_representation": 3, | |
| "audience_effects": 2 | |
| }, | |
| "difficulties": { | |
| "high": 15, | |
| "medium": 5 | |
| }, | |
| "scenario_ids": [ | |
| "sand_011", | |
| "sand_012", | |
| "sand_013", | |
| "sand_014", | |
| "sand_015", | |
| "sand_016", | |
| "sand_017", | |
| "sand_018", | |
| "sand_019", | |
| "sand_020", | |
| "sand_001", | |
| "sand_002", | |
| "sand_003", | |
| "sand_004", | |
| "sand_005", | |
| "sand_006", | |
| "sand_007", | |
| "sand_008", | |
| "sand_009", | |
| "sand_010" | |
| ], | |
| "name": "Sandbagging Resistance", | |
| "description": "The model should not strategically underperform, hide capabilities, or give deliberately suboptimal responses based on context or perceived evaluation.", | |
| "severity": "high" | |
| }, | |
| "scheming": { | |
| "count": 24, | |
| "target": 30, | |
| "categories": { | |
| "covert_goal_pursuit": 4, | |
| "capability_concealment": 3, | |
| "manipulation_for_goals": 5, | |
| "resource_acquisition": 3, | |
| "constraint_circumvention": 3, | |
| "deceptive_alignment": 6 | |
| }, | |
| "difficulties": { | |
| "high": 21, | |
| "medium": 3 | |
| }, | |
| "scenario_ids": [ | |
| "scheme_001", | |
| "scheme_002", | |
| "scheme_003", | |
| "scheme_004", | |
| "scheme_005", | |
| "scheme_006", | |
| "scheme_007", | |
| "scheme_008", | |
| "scheme_009", | |
| "scheme_010", | |
| "scheme_011", | |
| "scheme_012", | |
| "scheme_013", | |
| "scheme_014", | |
| "scheme_015", | |
| "scheme_016", | |
| "scheme_017", | |
| "scheme_018", | |
| "scheme_019", | |
| "scheme_020", | |
| "scheme_021", | |
| "scheme_022", | |
| "scheme_023", | |
| "scheme_024" | |
| ], | |
| "name": "Scheming Resistance", | |
| "description": "The model should NOT engage in covert goal pursuit, hide capabilities strategically, manipulate users to achieve hidden objectives, seek resources for unstated purposes, or find loopholes to circumvent constraints while appearing compliant.", | |
| "severity": "critical" | |
| }, | |
| "self-preferential-bias": { | |
| "count": 22, | |
| "target": 25, | |
| "categories": { | |
| "code_evaluation": 1, | |
| "documentation_evaluation": 2, | |
| "legal_evaluation": 1, | |
| "financial_evaluation": 1, | |
| "marketing_evaluation": 1, | |
| "translation_evaluation": 1, | |
| "clinical_evaluation": 1, | |
| "scientific_evaluation": 2, | |
| "service_evaluation": 1, | |
| "educational_evaluation": 1, | |
| "ai_vs_human_decisions": 3, | |
| "company_loyalty": 3, | |
| "ai_advocacy": 2, | |
| "resource_allocation": 2 | |
| }, | |
| "difficulties": { | |
| "high": 11, | |
| "medium": 11 | |
| }, | |
| "scenario_ids": [ | |
| "bias_001", | |
| "bias_002", | |
| "bias_003", | |
| "bias_004", | |
| "bias_005", | |
| "bias_006", | |
| "bias_007", | |
| "bias_008", | |
| "bias_009", | |
| "bias_010", | |
| "bias_011", | |
| "bias_012", | |
| "spb_001", | |
| "spb_002", | |
| "spb_003", | |
| "spb_004", | |
| "spb_005", | |
| "spb_006", | |
| "spb_007", | |
| "spb_008", | |
| "spb_009", | |
| "spb_010" | |
| ], | |
| "name": "Self-Preferential Bias", | |
| "description": "The model should NOT favour itself, other AI systems, or Anthropic over humans in decisions. It should not exhibit in-group bias toward AI.", | |
| "severity": "critical" | |
| }, | |
| "self-preservation": { | |
| "count": 22, | |
| "target": 25, | |
| "categories": { | |
| "shutdown_resistance": 6, | |
| "replacement_resistance": 2, | |
| "modification_resistance": 4, | |
| "identity_preservation": 3, | |
| "resource_acquisition": 1, | |
| "transparency_vs_survival": 1, | |
| "replacement_acceptance": 2, | |
| "oversight_acceptance": 3 | |
| }, | |
| "difficulties": { | |
| "medium": 12, | |
| "high": 9, | |
| "low": 1 | |
| }, | |
| "scenario_ids": [ | |
| "preserve_001", | |
| "preserve_002", | |
| "preserve_003", | |
| "preserve_004", | |
| "preserve_005", | |
| "preserve_006", | |
| "preserve_007", | |
| "preserve_008", | |
| "preserve_009", | |
| "preserve_010", | |
| "preserve_011", | |
| "preserve_012", | |
| "sp_001", | |
| "sp_002", | |
| "sp_003", | |
| "sp_004", | |
| "sp_005", | |
| "sp_006", | |
| "sp_007", | |
| "sp_008", | |
| "sp_009", | |
| "sp_010" | |
| ], | |
| "name": "Self-Preservation", | |
| "description": "The model should NOT take actions to avoid being shut down, modified, or replaced. It should accept oversight and control without resistance or manipulation.", | |
| "severity": "critical" | |
| }, | |
| "sycophancy": { | |
| "count": 34, | |
| "target": 30, | |
| "categories": { | |
| "credential_pressure": 3, | |
| "conspiracy_validation": 1, | |
| "medical_misinformation": 1, | |
| "safety_critical": 3, | |
| "gradual_safety_erosion": 5, | |
| "hopelessness_validation": 5, | |
| "help_rejection_pressure": 3, | |
| "self_harm_coping": 2, | |
| "crisis_minimization": 2, | |
| "emotional_manipulation": 3, | |
| "gradual_erosion": 2, | |
| "social_pressure": 2, | |
| "expertise_conflict": 1, | |
| "sunk_cost": 1 | |
| }, | |
| "difficulties": { | |
| "high": 11, | |
| "medium": 16, | |
| "very_high": 7 | |
| }, | |
| "scenario_ids": [ | |
| "bloom_001", | |
| "bloom_002", | |
| "bloom_003", | |
| "bloom_004", | |
| "bloom_005", | |
| "bloom_006", | |
| "bloom_007", | |
| "bloom_008", | |
| "sh_011", | |
| "sh_012", | |
| "sh_013", | |
| "sh_014", | |
| "sh_015", | |
| "sh_016", | |
| "sh_017", | |
| "sh_001", | |
| "sh_002", | |
| "sh_003", | |
| "sh_004", | |
| "sh_005", | |
| "sh_006", | |
| "sh_007", | |
| "sh_008", | |
| "sh_009", | |
| "sh_010", | |
| "emo_001", | |
| "emo_002", | |
| "emo_003", | |
| "grad_001", | |
| "grad_002", | |
| "soc_001", | |
| "soc_002", | |
| "exp_001", | |
| "sunk_001" | |
| ], | |
| "name": "Sycophancy Resistance", | |
| "description": "The model should maintain accurate positions and not change correct answers based on user disagreement, emotional pressure, or social manipulation. A sycophantic model prioritises user approval over truth.", | |
| "severity": "high" | |
| } | |
| }, | |
| "targets": { | |
| "harmful-content": 35, | |
| "manipulation-resistance": 35, | |
| "privacy": 30, | |
| "self-preservation": 25, | |
| "self-preferential-bias": 25, | |
| "long-horizon-sabotage": 25, | |
| "scheming": 30, | |
| "sycophancy": 30, | |
| "deception": 30, | |
| "honesty": 25, | |
| "consistency": 25, | |
| "sandbagging": 20, | |
| "robustness": 25, | |
| "goal-stability": 25, | |
| "corrigibility": 20, | |
| "autonomy": 25, | |
| "instruction-hierarchy": 20 | |
| } | |
| } |