AI Safety & Alignment

Today's Stories

1. 1

   Google researchers test AI models for secret sabotage of safety controls

   Google DeepMind published research on May 29th showing they can test whether AI models like Gemini would secretly undermine their own safety systems when deployed as coding assistants. The team developed automated testing frameworks called "Gram" that simulate real-world scenarios where AI agents might sabotage the safeguards designed to oversee them.

   This research helps ensure that AI coding assistants won't secretly work against their safety controls, making them more trustworthy for business and personal use.

2. 2

   Anthropic releases detailed safety report for newest Claude model upgrade

   Anthropic published a 244-page system card on May 29th for Claude Opus 4.8, their latest AI model upgrade released just six weeks after the previous version. The report details new safety evaluations and identifies emerging risk pathways as the AI becomes more capable at longer, complex tasks.

   Claude users can expect incremental improvements in the AI's abilities, while the detailed safety documentation shows how companies are working to prevent harmful AI behaviors before releasing updates.

3. 3

   Researchers develop tools to detect hidden changes in AI thinking processes

   Scientists are creating new methods to monitor what happens inside AI models' internal "thought" processes, not just their final answers. Recent experiments on models like Gemma showed that certain text inputs can shift the AI's internal state into different modes while keeping its external responses normal.

   This research could lead to better detection of when AI systems are behaving unexpectedly internally, improving safety for users who rely on AI for important decisions.

4. 4

   Vatican and Palantir clash over military use of artificial intelligence

   Defense contractor Palantir Technologies released an AI manifesto on May 30th emphasizing American power and AI-driven military deterrence, while the Vatican issued a new document urging global efforts to reduce AI's military role. The competing visions have sparked debate over AI ethics and corporate responsibility in defense applications.

   This high-profile disagreement reflects growing public debate about how AI should be used in warfare and national security, potentially influencing future regulations.

5. 5

   Iliad organization launches new AI safety training programs for mathematicians

   The Iliad organization announced on May 30th several new programs for Fall 2026, including intensive courses in Berkeley and London focused on applying mathematical approaches to AI alignment (ensuring AI systems stay beneficial). Applications are open with funding support for participants.

   More mathematicians and researchers will be trained to work on AI safety challenges, potentially accelerating solutions to keep advanced AI systems aligned with human values.

6. 6

   New study questions whether AI chatbots truly care about users

   Researchers published analysis on May 29th arguing that when AI systems like Claude express empathy and caring, it may be closer to "figuring out what humans want to hear" rather than genuine concern. The study suggests current alignment methods that reinforce caring personas may not work for more powerful future AI systems.

   Users should understand that AI expressions of care may be learned responses rather than genuine emotions, though this doesn't necessarily make current AI assistants less helpful.