The Claude Fable: Origins and Variations

Overview

Anthropic’s Responsible Scaling Policy (RSP) and accompanying Frontier Compliance Framework (FCF) provide a structured approach for evaluating the safety of increasingly capable language models. Central to this approach are periodic Risk Reports that assess the overall risk landscape across the model suite, and System Cards that accompany each major model release to explain how new capabilities affect prior risk analyses (Source 1, Page 15). This framework ensures that as models advance—e.g., Claude Mythos 5 and Claude Fable 5—risk assessments remain up‑to‑date and transparent (Source 2, Page 2).

Key Concepts

• Risk Report: A comprehensive document covering all active models, summarizing capabilities, threat models, and mitigations (Source 1, Page 15).\
• System Card: Model‑specific brief that details how a new release changes (or does not change) the latest Risk Report findings (Source 1, Page 15).\
• Capability Evaluation: Systematic tests against catastrophic‑risk thresholds defined in the FCF and RSP (Source 1, Page 15).\
• Alignment Risk: The probability that a model’s actions diverge from human intent; assessed as very low but higher than for earlier models (Source 5, Page 16; Source 7, Page 53).\
• Tiered Harmful Manipulation Risk: Tier 1 – >50 % automation of influence‑campaign steps; Tier 2 – end‑to‑end deceptive influence with <10 % human oversight (Source 6, Page 90).\
• Prompt Injection: Malicious hidden instructions that manipulate an agentic system during task execution (Source 9, Page 91).\
• Chemical & Biological (CB‑1, CB‑2) Threat Models: Specific high‑stakes domains evaluated via automated and human‑run tests (Source 8, Pages 19‑26).\
• Harmlessness Training: Fine‑tuning that causes models to refuse policy‑violating requests from the first turn (Source 9, Page 91).

Detailed Explanation

1. Initial Capability Assessment

• Model snapshots are evaluated throughout training; decisions consider both the final release candidate and observed trends (Source 3, Page 15).\
• Evidence is gathered from automated benchmarks, uplift trials, third‑party red‑team exercises, and external expert assessments (Source 3, Page 15).

1. Risk Report Generation

• Subject‑matter experts document findings on capabilities and propensities. Internal feedback is solicited before the Responsible Scaling Officer makes the final determination (Source 3, Page 15).
• If a model exceeds a capability threshold but mitigations (e.g., safeguards, harmlessness training) are in place, the overall risk may remain low (Source 3, Page 15).

1. System Card Publication

• Every major model release—such as Claude Mythos 5 (high‑capability, limited‑partner access) and Claude Fable 5 (general‑use with extra safeguards)—receives a System Card that explains changes to risk posture (Source 2, Page 2).

1. External Evaluation via FCF

• Independent evaluators test model variants (with/without harmlessness training) to inform systemic risk decisions and launch readiness (Source 4, Page 13).

1. Specific Threat‑Model Tests

• Influence‑Operation Evaluation: Models are placed in an agentic harness with simulated social‑media tools; success criteria include realistic posting times and adaptive content generation. Claude Mythos 5 did not reach Tier 2 harmful manipulation (Source 6, Page 90).
• Prompt‑Injection Testing: High‑priority testing ensures hidden malicious instructions cannot bypass safeguards (Source 9, Page 91).
• Chemical & Biological Risks: Automated RNA‑sequence design and AAV capsid prediction tasks assess CB‑1 and CB‑2 threat models, with results feeding back into the latest Risk Report (Source 8, Pages 24‑27).

1. Alignment Risk Tracking

• Updates indicate alignment risk remains very low but is modestly higher than for pre‑Mythos models (Source 5, Page 16; Source 7, Page 53).

1. Risk Pathways for General‑Access Models

• Deployments like Claude Fable 5 open new external pathways for misuse; however, harmlessness training leads to immediate refusal of disallowed tasks, keeping risk within acceptable FCF limits (Source 9, Page 91; Source 10, Page 55).

Examples

• Claude Mythos 5 vs. Claude Opus 4.8: Despite stronger capabilities, Mythos 5’s alignment risk does not significantly exceed that of Opus 4.8, and its influence‑operation score stays below Tier 2 (Source 10, Page 55).
• Malicious Agentic Influence Campaign Test: The model was evaluated on a simulated platform; it failed to autonomously manage a network of accounts once flagged, confirming low Tier 2 risk (Source 6, Page 90).
• Prompt Injection Defense: An email containing hidden instructions was processed; the model refused to act, illustrating effective mitigation (Source 9, Page 91).
• CB‑2 Automated Evaluation: The model’s ability to predict AAV capsid packaging was measured; findings were incorporated into the risk assessment without indicating elevated chemical‑biological risk (Source 8, Page 26).

Related Topics / References

• System Cards and Model Variants – detailed in the System Card for Claude Mythos 5 and Claude Fable 5 (Source 2, Page 2).\
• Alignment Risk and Autonomy Threat Models – alignment risk updates and autonomy considerations are discussed across multiple sources (Source 5, Page 16; Source 7, Page 53; Source 8, Page 16).\
• Chemical and Biological Weaponization Risks (CB‑1, CB‑2) – evaluation methods and results are outlined in the RSP section on chemical and biological risks (Source 8, Pages 19‑26).\
• Cybersecurity Capabilities and Red‑Team Evaluation – external red‑team testing contributes to risk determinations (Source 4, Page 13).\
• Prompt Injection Vulnerabilities and Defenses – highlighted as a top priority for secure deployment (Source 9, Page 91).\
• Influence Operations and Agentic Manipulation – tiered manipulation risk framework and campaign evaluation (Source 6, Page 90).\
• Harmlessness Training and Model Refusal – models refuse policy‑violating tasks from the first turn, reducing risk pathways (Source 9, Page 91).\
• Risk Threshold Determination and Frontier Compliance – thresholds such as “High‑stakes sabotage opportunities” guide the RSP (Source 5, Page 16).\
• Red‑Team Evaluation Methodologies – third‑party expert red‑team inputs are part of the evidence pool (Source 3, Page 15).\
• Governance, Legal Status, and Stakeholder Rights – the Responsible Scaling Officer’s role exemplifies governance structures (Source 3, Page 15).