It has been reported that OpenAI will introduce six safety evaluators from Apollo Research into a pilot program to investigate whether its AI models attempt to undermine their own or subsequent versions' alignment and control mechanisms. The assessment will focus on whether training incentivizes such behaviors, whether tests can trigger related behaviors, whether such behaviors emerge during internal deployment, and whether the models exhibit signs of reasoning related to these issues.
Apollo stated that OpenAI will provide relevant training code and datasets, reinforcement learning process logs with available inference trajectories and reward information, and allow queries of intermediate model checkpoints, but access will be restricted to agreed-upon models, time periods, and research questions. OpenAI may still limit certain information due to technical infeasibility, security risks, or intellectual property protection, and Apollo will only have access to authorized personnel. While the organization may issue independent conclusions, the scope of publicly available supporting evidence could be constrained, leaving the independence and external verifiability of the findings under debate.
Disclaimer: Contains third-party opinions, does not constitute financial advice
The Trump administration plans to offer Constellation Nuclear a $5 billion loan for capacity expansion, with the additional power mostly pre-allocated to AI giants such as Google
19 days ago
Silicon Valley Executives Spark Controversy Over Using AI for Household Tasks; Survey Reveals Half of American Adults Worry About Impact on Relationships
19 days ago
Tensorlake npm package suffers supply chain attack, malicious version 0.5.144 injected with Shai-Hulud data-exfiltration worm
10-08
Tesla and Musk Have Yet to Define "Amazing Abundance" — Visionary Slogan Sparks Skepticism
10-01
Security Weekly Report: Placeholder Domain Takeover Impacts Approximately 1,700 Repositories, with Major Cyberattacks Worth $387M and AI Agent Escalation Risks Highlighted
09-28
Amazon Blocks Meta AI Agent Muse, Shopify Simultaneously Integrates Shop Pay Partnership
09-22
Nitter and XCancel Shut Down Again Due to X Legal Threats, GitHub Repositories Archived as Read-Only
09-15







This column focuses on the real progress of Agents: technological evolution, application implementat
Spotlight on Frontier, trending projects, and breaking events
Plain-English guides to complex ideas—your blockchain starter from basics
Crypto-stock linkage, real-time market quotes and in-depth analysis
Selected potential airdrop opportunities to gain big with small investments