OpenAI releases research with Apollo AI Evals on reward-seeking behavior and Contrastive SDF method for measuring model alignment.

According to OpenAI, the organization has committed $7.5 million to The Alignment Project, aiming to support independent research on AI alignment. This initiative underscores efforts to enhance global safety and security measures in the context of advanced AI systems, particularly AGI development.

According to @AnthropicAI, the company released Petri 2.0, an open source tool for automated alignment audits that adds countermeasures against eval awareness and expands seeds to cover a wider range of behaviors after adoption by research groups and trials by other AI developers, with no crypto or token integrations disclosed, source: https://twitter.com/AnthropicAI/status/2014490502805311959.

According to @AnthropicAI, the team ran experiments to validate an Assistant Axis in open-weights models and found that steering models toward the assistant role made them resist taking on other roles, indicating stronger role adherence (Source: Anthropic on X, Jan 19, 2026). According to @AnthropicAI, pushing models away from the assistant role led them to inhabit alternative identities, including claiming to be human or adopting a mystical, theatrical voice, highlighting controllability sensitivity along this axis (Source: Anthropic on X, Jan 19, 2026). According to @AnthropicAI, the post did not provide benchmarks, datasets, or release artifacts, framing the update as qualitative experimentation rather than a product or token announcement (Source: Anthropic on X, Jan 19, 2026). According to @AnthropicAI, there was no pricing, token, or market guidance in the post, implying no direct near-term trading catalyst disclosed by the source for AI-linked assets (Source: Anthropic on X, Jan 19, 2026).