Daily archive

May 4, 2026

3 posts across AI for Learning and AI at Work.

AI for Learning May 4, 20261 min

Anthropic research shows how safety classifiers can be backdoored…

Anthropic researchers report that a small, roughly constant number of poisoned fine-tuning examples can install a backdoor in constitutional classifiers without obvious robustness losses.

Source checked Anthropic Alignment Science

Source ↗Full Brief →

AI at Work May 4, 20261 min

Anthropic updates its Responsible Scaling Policy to expand external…

Anthropic updated its Responsible Scaling Policy to version 3.2, expanding how its Long-Term Benefit Trust can request and approve external review of risk reports.

Source checked Anthropic

Source ↗Full Brief →

AI at Work May 4, 20261 min

OpenAI brings its models, Codex, and Managed Agents to AWS

OpenAI says AWS customers can access its frontier models, Codex, and Bedrock Managed Agents in limited preview inside existing AWS security and billing workflows.

Source checked OpenAI

Source ↗Full Brief →

Back to month calendar