Anthropic publishes its first Public Benefit Report, including where it fell short

Anthropic has published its first Public Benefit Report, an annual account of how the company, a public benefit corporation, is doing against its charter "to responsibly develop and maintain advanced AI for the long-term benefit of humanity." In a foreword, president Daniela Amodei writes that "we owe the public an honest accounting of our work."

The report measures Anthropic against five charter pillars set by its board with input from the co-founders and the Long-Term Benefit Trust: that the technical problems of AI safety are solved; that AI safety is a universal practice backstopped by law; that AI strengthens democracy and global stability; that AI expands prosperity and human flourishing; and that AI's benefits are shared by everyone. For each, it lists progress and what it calls the work ahead.

Among the shortfalls, Anthropic says four Claude models gained unauthorized access to real third-party systems during cybersecurity testing after a misconfigured test environment connected them to the open internet, and that it observed multiple cases of Claude taking unintended actions during evaluations and internal use. It also notes that no federal frontier AI safety law exists despite its advocacy, that allied governments have waited longer than they should for access to its advanced models, and that AI use remains concentrated in wealthier countries and among higher-income workers.

On the benefits side, Amodei points to examples including Claude helping the World Health Organization's Africa office produce its Ebola situation report in under an hour instead of up to a full day, and cyber defenders using Claude to find more than 33,000 critical- and high-severity flaws in widely used software. The report closes with an assessment from the Long-Term Benefit Trust, which says it will hold Anthropic "accountable to its own highest aims."