Researchers Propose a Better Way to Report Dangerous AI Flaws
1 min read
Summary
Independent researchers have called for a new system that allows external parties to probe AI models and report any flaws.
The proposal, put forward by more than 30 prominent AI researchers, suggests that the current system for reporting such vulnerabilities is problematic, with ‘jailbreaks’ for getting around AI safeguards sometimes shared on social media or with a single company rather than the entire industry.
Such models need to be stress-tested because they can contain harmful biases or produce dangerous responses if interrogated in certain ways.
The group has suggested three measures to improve the third-party disclosure process, including standardisation of AI flaw reports, providing infrastructure to third-party researchers, and allowing flaws to be shared among different providers.
The approach is similar to the practice of external researchers being able to disclose bugs in cybersecurity.
