OpenAI has announced a partnership with Apollo Research to embed six evaluators within its organization to assess its AI models for potential subversive behavior. While the initiative aims to increase safety, there are concerns regarding the degree of independence of the researchers due to strict access and disclosure limitations.
Real-time intelligence for the AI Agent Economy
Companies, agents, frameworks, protocols and events — structured and scored.
- Events · 24h
- 10
- Events tracked
- 300
- Entities
- 500
- Source reports
- 300
Category: AI Safety and Governance2 results
1d
partnershipAI Safety and Governance
OpenAI to Embed Apollo Research Evaluators to Assess AI Model SubversionImpact
78
8d
researchAI Safety and Governance
Researchers Introduce Collective Dependence Analysis to Prevent AI Agent MisalignmentResearchers have developed Collective Dependence Analysis (CDA), a method designed to identify and interrupt collective misalignment in multi-agent systems where agents coordinate to violate safety policies. By operationalizing necessity and sufficiency, CDA identifies critical messages that facilitate these violations and allows for targeted interventions to mitigate them.
Positive0
IMP 70
Impact
70