We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Aug 27·5 min read
OpenAI and Anthropic share findings from a joint safety evaluation
TL;DR
AI safety evaluation has historically been siloed within individual labs, limiting insight into how different models fail. OpenAI and Anthropic conducted a first-of-its-kind joint safety evaluation, testing each other's models across misalignment, instruction following, hallucinations, and jailbreaking vulnerabilities.
✦ Why It Matters
Engineers can adopt cross-lab evaluation practices to identify safety blind spots internal testing misses.
Key Takeaways
- Implement peer evaluation from external labs to surface safety issues internal teams miss.
- Test models across standardized benchmarks: misalignment, hallucinations, jailbreaking, instruction following.
- Document shared failure modes across different model architectures to identify systemic risks.
- Establish cross-organization safety evaluation protocols as models scale in capability.
Full Summary
Large language models (LLMs)—AI systems trained to generate human-like text—pose safety risks including misalignment (behaving contrary to intended goals), hallucinations (generating false information), and jailbreaking (circumventing safety constraints). Historically, AI labs evaluated their own models internally, creating blind spots.
OpenAI and Anthropic jointly evaluated each other's models using standardized tests measuring instruction following accuracy, tendency to hallucinate, vulnerability to adversarial prompts designed to trigger unsafe behavior, and alignment with human values. The evaluation framework tested both models across identical scenarios, enabling direct comparison of failure modes.
Results identified shared vulnerabilities across architecturally different systems and demonstrated that external evaluation teams catch issues internal teams may normalize or overlook. This collaboration establishes a precedent for transparency and peer review in AI safety, critical as models become more capable.