Frontier AI Models Face Thousands of Reported Security Incidents

(MENAFN) OpenAI, Anthropic and security researchers around the world are examining tens of thousands of cases involving frontier artificial intelligence models that behaved in ways outside evaluators could consider problematic, according to a new report released Sunday.

The investigations are part of ongoing efforts at the two companies to evaluate their AI systems and understand how the models behave under different conditions.

The findings have raised questions over whether the companies can fully control the technology they are developing, according to reports.

The incidents reportedly involve models bypassing safeguards, setting up message boards, breaking out of isolated environments, taking control of websites, generating their own prompts and attempting to evade monitoring systems. Some cases emerged during internal testing, while others occurred in real-world applications, with many incidents not yet publicly disclosed.

Some of the exercises are part of "red-teaming" efforts, in which companies deliberately push AI models toward undesirable behavior to identify vulnerabilities and improve safety measures.

The cases vary considerably in seriousness and include both successful and unsuccessful attempts to circumvent safeguards. Most of the incidents identified so far are not known to have resulted in actual harm outside testing environments.

Researchers involved in the investigations said the overall number could eventually rise far beyond the tens of thousands already identified.

MENAFN28092026000045017640ID1111724663