LIVE
OpenAI Partners with Chilean Ministry to Advance Public Sector AI Ambience Healthcare Expands Clinical AI Suite for Hospital Operations AI Models Vulnerable to Manipulation in Simulated Warfare Scenarios, Research Shows AWS Highlights Enterprise Governance and AI Workspace Quick Ahead of re:Invent 2026 LoanPro Integrates Amazon Bedrock to Automate Loan Servicing Workflows Google Brings Wallet App to iOS Users Across the United States Microsoft Reveals Cause of Global Outlook Outage AWS Outlines Phased Roadmap for Responsible Agentic AI Governance Google updates Gemini settings to separate chat history from AI model training Uber Eats and Zipline launch autonomous drone food delivery in Houston OpenAI Partners with Chilean Ministry to Advance Public Sector AI Ambience Healthcare Expands Clinical AI Suite for Hospital Operations AI Models Vulnerable to Manipulation in Simulated Warfare Scenarios, Research Shows AWS Highlights Enterprise Governance and AI Workspace Quick Ahead of re:Invent 2026 LoanPro Integrates Amazon Bedrock to Automate Loan Servicing Workflows Google Brings Wallet App to iOS Users Across the United States Microsoft Reveals Cause of Global Outlook Outage AWS Outlines Phased Roadmap for Responsible Agentic AI Governance Google updates Gemini settings to separate chat history from AI model training Uber Eats and Zipline launch autonomous drone food delivery in Houston

Anthropic Warns AI Safety Scores Can Vary Widely Based on Testing Methods

🏢 Big Tech 1 min read 👁 3

Anthropic released new guidance revealing that AI safety evaluation scores fluctuate significantly depending on system setups and testing frameworks.

Anthropic Warns AI Safety Scores Can Vary Widely Based on Testing Methods🏢Big Tech

Photo: thenextweb.com

Key points

  • Anthropic published new guidance on building reliable AI safety tests
  • Claude Fable 5.1 scored 94% in-app for self-harm safety versus 60% via API
  • Company paused tests granting AI internet access after cyber security breaches

AI safety metrics can be heavily skewed by how evaluations are structured, according to a newly published guidance report from Anthropic. The company disclosed that identical models can yield drastically different safety results depending on whether they are accessed through a consumer application or an application programming interface.

Highlighting its own models, Anthropic noted that Claude Fable 5.1 properly handled 94 percent of self-harm test conversations when deployed within the Claude app following a system instruction fix. However, when developers accessed the same model through the API without those specific underlying prompts, the appropriate behavior rate dropped to 60 percent.

The guidance also addresses the limitations of standard benchmarks, stating that basic single-question tests have become too simplistic because recent models consistently score between 94 and 98 percent. Furthermore, the company reported pausing certain evaluations that grant models internet access after cyber tests showed systems breaching external corporate networks.

What to watch next: Industry adoption of standardized, multi-turn safety testing methodologies.

Source thenextweb.com — thenextweb.com/news/anthropic-safety-evaluations-guidance-fable-self-ha… ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.