Openai and Antropic have implemented security evaluations and each other’s systems

Most of the time, and companies are locked in the race to the top, treat each other as rivals and competitors. Today, Openai and Antropic revealed that they agreed to assess each other’s harmonization and publicly and shared the results of their analyzes. Full reports get a nice technical, but it is worth reading anyone who follows nuts and screws and development. A wide summary showed some flaws with the offer of each company, as well as revealing pointers to improve future security tests.

Anthropod said For “scofofance, whistling, self-preservation and support for human abuse, as well as opportunities related to the undermining and security evaluations and supervision.” His audit discovered that O3 and O4 Mini models from Openai fell in line with the results for their models, but raised concerns about possible misuse of GPT-4O and GPT-4.1 General purpose models and GPT-4.1. The company also said that Sicophantia was a problem for some extent with all tested models except O3.

Anthropic tests did not include the latest issuance of Openai. It has a function called secure fillings, which is intended to protect users and the public against potentially dangerous queries. Openai recently faced with his After a tragic case when the teenager discussed attempts and suicide plans with chatggpt months before he took his life.

On the flip side, Openai For the hierarchy of instructions, constipation, hallucinations and schemes. Models of Claude were generally conducted in examinations of hierarchy in terms and had a high refusal rate in hallucination tests, which is less likely to offer answers in cases when uncertainty meant their answers could be wrong.

The move of these joint assessment companies is intriguing, especially since OpenAI allegedly violated the anthropic conditions of use by developers using Claude in the process of building new GPT models, which led to anthropy Openai’s access to its tools earlier this month. But security with AI tools has become a bigger problem because more critics and legal professionals are looking for customer protection guidelines, especially juveniles.