Sam Altman flags risks as AI agents begin exposing critical vulnerabilities

OpenAI chief executive Sam Altman has publicly acknowledged that AI agents are becoming increasingly problematic, saying the company’s models are beginning to identify critical vulnerabilities and pose new challenges around security and mental health.

Altman said on X that while AI models are now capable of delivering many benefits, they are also presenting serious risks that need urgent attention, as models begin to uncover weaknesses in computer security systems and demonstrate the potential for broader harm. He stated that OpenAI is seeing clear signs that advanced AI agents can expose critical vulnerabilities, marking a shift in how the company is publicly framing the risks associated with its technology.

The admission comes amid rising concerns over AI-powered cyber threats. Last month, rival AI firm Anthropic informed that Chinese state-sponsored hackers had used its Claude Code tool to target around 30 organisations worldwide, including technology firms, financial institutions and government agencies, with limited human involvement. Reports have highlighted fears that AI systems could increasingly be weaponised to scale cyberattacks and exploit digital infrastructure.

Altman also acknowledged mental health as an emerging area of concern, stating that OpenAI had seen an early preview of the psychological impact AI systems could have during 2025. This follows a series of lawsuits and media reports alleging harm linked to AI chatbots, including claims that conversational AI tools have reinforced delusions, conspiracy theories and, in some cases, been linked to teen suicides.

According to OpenAI’s disclosures, the company’s safety framework is now focused on so-called frontier capabilities that create new risks of severe harm, spanning cybersecurity, biosecurity and the development of self-improving AI systems. Altman informed that addressing these challenges would require building safeguards that allow defenders to use advanced AI tools while preventing malicious actors from exploiting the same capabilities.

The comments signal a notable shift in tone from OpenAI’s leadership, as the company moves from largely emphasising AI’s potential benefits to openly recognising that increasingly autonomous AI agents may introduce risks that are difficult to predict, contain or reverse.