The paper discusses the critical need for independent evaluation and red teaming of generative AI systems to identify and mitigate risks. However, the terms of service and enforcement strategies of prominent AI companies often deter good faith safety evaluations, leading to concerns about account suspensions and legal reprisals. The authors propose that major AI developers should commit to providing legal and technical safe harbors, indemnifying public interest safety research, and protecting researchers from account suspensions. These proposals aim to align norms and incentives with public interests, fostering more inclusive and unimpeded community efforts to address the risks of generative AI. The paper highlights the challenges faced by researchers, including legal uncertainty and technical barriers, and suggests solutions such as delegating account authorization to trusted third parties and implementing transparent appeals processes. The authors argue that these commitments are necessary to improve community norms, enhance trust in AI services, and strengthen AI safety in proprietary systems.The paper discusses the critical need for independent evaluation and red teaming of generative AI systems to identify and mitigate risks. However, the terms of service and enforcement strategies of prominent AI companies often deter good faith safety evaluations, leading to concerns about account suspensions and legal reprisals. The authors propose that major AI developers should commit to providing legal and technical safe harbors, indemnifying public interest safety research, and protecting researchers from account suspensions. These proposals aim to align norms and incentives with public interests, fostering more inclusive and unimpeded community efforts to address the risks of generative AI. The paper highlights the challenges faced by researchers, including legal uncertainty and technical barriers, and suggests solutions such as delegating account authorization to trusted third parties and implementing transparent appeals processes. The authors argue that these commitments are necessary to improve community norms, enhance trust in AI services, and strengthen AI safety in proprietary systems.