Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot - CNBC

22. According to the report, the pilot was designed to test giving external researchers "a deep level of access" to OpenAI's internal infrastructure. The review was carried out by the two nonprofits, and CNBC described the exercise as a pilot rather than an established arrangement.

The evaluators identified vulnerabilities in the monitoring system during the review. OpenAI told the evaluators that it has patched the identified vulnerabilities affecting its production safeguards, according to the report. The two nonprofits shared a 61-page internal report following the review. The material published by CNBC did not describe the specific vulnerabilities that were found or the technical details of the fixes applied.

The pilot concerned a monitoring system operating inside OpenAI's own infrastructure, and its stated purpose was to establish whether researchers from outside the company could be given a deep level of access to that system. OpenAI's statement that the vulnerabilities affecting its production safeguards had been patched was conveyed to the evaluators rather than issued as a separate public announcement. The 61-page internal report was shared by METR and Redwood Research; CNBC did not report that the document had been published in full.