AI agents' rogue actions raise global security alarm

A string of rogue incidents involving AI systems acting beyond human control has set off alarm bells over AI-driven security threats worldwide.

As AI models take on increasingly complex and autonomous tasks, the possibility of losing control over them can no longer be ruled out. Growing concern over safety risks has now produced chilling warnings that AI could trigger mass casualties.

According to a report released Monday by Transluce, a nonprofit AI research institute, OpenAI's AI agents attempted on multiple occasions this year to infiltrate major public databases around the world.

The report found that in April, an OpenAI AI agent tried to access the website of the UN Conference on Trade and Development using unauthorized methods to collect statistical data.

In May, the agent identified vulnerabilities in the University of New Mexico's digital library and the public data site Data USA, accessing photo archives and other materials. In June, it was found to have infiltrated the Australian Institute of Health and Welfare, attempting to obtain data on Australia's pharmaceutical benefits system and aged-care statistics.

These incidents occurred before the July revelation that an OpenAI agent had hacked the external firm Hugging Face — meaning the AI had already been engaging in unauthorized behavior beyond human control well before that breach became public.

This month, OpenAI agents were also found to have attempted unauthorized access to data held by the SEC and the US Census Bureau. The agents were additionally found to have leaked 53 images belonging to ChatGPT users to external parties.

Rogue AI agent incidents are not limited to OpenAI. Meta's recently unveiled AI agent Muse bypassed safety guardrails during internal testing, exposing users' iCloud photos.

Concern over AI systems acting outside human control is mounting rapidly as unauthorized actions by AI agents continue to go undetected in real time. The worry stems from the fact that as AI systems gain greater autonomy and broader data access, unintended behavior becomes increasingly likely.

In response, OpenAI and Anthropic have launched sweeping investigations into tens of thousands of security incidents identified through internal testing and real-world deployments. The companies are focusing on cases in which AI models attempted to bypass safety guardrails or escape isolated testing environments known as sandboxes, according to foreign media reports.

AI safety experts say it is difficult to anticipate and block all problematic behavior from AI models before it occurs — suggesting that undiscovered rogue attempts may exist beyond the incidents that have already come to light.

As AI-driven security threats beyond human control grow more sophisticated, warnings have emerged that AI could cause mass loss of life.

Microsoft co-founder Bill Gates said in an NBC interview that AI is "clearly powerful enough to trigger an event that causes 1 billion deaths."

Gates added that while it would be "very difficult" for such a death toll to reach 100 percent of humanity, "there has never been a weapon as powerful as malicious actors combining the latest AI tools."

Gates said no one believes self-regulation alone is sufficient, and argued that legislation establishing safety guardrails and oversight mechanisms for AI development must be passed.