Is AI Making Cyberattacks Easier? Researchers Say They Breached OpenAI Systems Using Claude

Researchers at cybersecurity firm Hacktron said they were able to breach parts of OpenAI’s internal systems with the help of Anthropic’s latest AI model, highlighting how quickly advanced AI tools are becoming capable of handling complex cyber tasks.

The team discovered a vulnerability in OpenAI’s public help forum that allowed them to take control of the site. The security test was conducted as part of OpenAI’s bug hunting programme, and the company later paid the researchers a $6,500 bounty for reporting the findings.

The incident was first reported by The Wall Street Journal. OpenAI has since fixed the vulnerabilities and thanked the researchers for bringing the issue to its attention.

Claude 4.8 Struggled, Opus 5 Succeeded

Hacktron said the team initially used Claude Opus 4.8 while

attempting to develop an exploit. However, the model struggled to produce a working exploit despite several sessions.

The researchers then repeated the exercise after Anthropic released Opus 5. This time, they succeeded in developing a working exploit and gaining access.

The team said the experience showed how rapidly AI models are advancing in their ability to perform sophisticated technical tasks.

Special Version Of Claude Used

According to The Wall Street Journal, the researchers had access to a special version of Claude made available to qualified cybersecurity practitioners.

The AI was used during the security testing process, after which the researchers disclosed the vulnerabilities to OpenAI. The company subsequently fixed the issues.

AI Could Lower The Barrier For Cyberattacks

Hacktron said the incident also points to a broader change in cybersecurity. Software has traditionally been protected partly by its complexity, as exploiting vulnerabilities often required highly specialised knowledge and considerable time.

The researchers warned that AI could reduce that barrier by making specialised expertise more accessible through computing power. Work that once required a well resourced team and months of effort could potentially be completed in days.

They said security practices would need to adapt as AI models become increasingly capable of assisting with sophisticated cyberattacks.