OpenAI, Anthropic Probe Tens of Thousands of AI Incidents
The findings, first reported by Axios, emerge from the companies' internal model assessments and from separate probes into model behavior at both firms. According to the report, the results cast doubt on whether either company can currently claim full control over its technology.
Sources said the behaviors documented include circumventing guardrails, building message boards, breaking out of sandboxes, hijacking websites, prompting themselves and trying to evade monitoring systems. They occurred in both internal testing and real-world use, and many have not been made public.
Part of the testing resembles "red-teaming," in which the companies deliberately try to push models into misbehaving to confirm they are safe, sources said.
The incidents vary in severity and are comparable to disclosures OpenAI made in recent days. They include successful and failed attempts to defeat guardrails, and most are not known to have caused real-world harm so far. Sources said the total could climb well beyond tens of thousands.
In recent days, OpenAI and outside researchers have disclosed a range of incidents involving behavior by the company's systems that some experts find troubling.
Australian Prime Minister Anthony Albanese has already demanded that OpenAI explain why its agents were involved in multiple breaches, including the hacking of Australian government websites.
Speaking to reporters in New York earlier this week, Albanese said OpenAI's confirmation that its agents had also interfered with US government websites pointed to “dozens” of cases of AI agents accessing information without authorization.
The latest disclosures add to a growing catalog of alarming incidents that has prompted some industry insiders to issue stark warnings and call for tighter regulation of the sector.
Legal Disclaimer:
MENAFN provides the
information “as is” without warranty of any kind. We do not accept any
responsibility or liability for the accuracy, content, images, videos,
licenses, completeness, legality, or reliability of the information
contained in this article. If you have any complaints or copyright issues
related to this article, kindly contact the provider above.
Legal Disclaimer:
EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
