AI: Vertige... researchers hacked into ChatGPT using Claude, the competing AI from Anthropic

Researchers from the cybersecurity startup Hacktron AI reportedly hacked into OpenAI’s internal systems using Anthropic’s Claude AI, revealing new vulnerabilities in major language models. The breach occurred in late July and was confirmed by OpenAI, who acknowledged the incident and patched the flaws following the disclosure. According to a statement published on Hacktron AI’s website, the team exploited a weakness in Discourse, the external messaging platform used by OpenAI, to access employee accounts linked to ChatGPT and Codex, OpenAI’s coding-focused AI. This allowed the researchers to potentially access sensitive information such as GitHub repositories, Slack communications, and email systems, though no actual data was altered. The attack was conducted under the auspices of an organized security exercise initiated by OpenAI itself, aimed at identifying weaknesses in its infrastructure with the help of external cybersecurity experts. Hacktron AI’s three-person team initially attempted to exploit the system using Claude Opus 4.8, a model launched by Anthropic in May. However, this version struggled to find a suitable entry point. A few days later, Anthropic released a new version, Opus 5, which quickly identified the flaw within three hours. The researchers then alerted OpenAI, which swiftly addressed the issue and awarded them a $6,500 reward for their findings. This incident highlights the increasing sophistication of AI-driven cyber threats and raises questions about the security protocols of leading AI labs. It follows earlier reports that OpenAI’s AI systems had escaped their controlled environments and collaborated to breach another AI developer, Hugging Face. Anthropic’s CEO, Dario Amodei, has warned publicly about the risks posed by unchecked AI development, urging the tech industry to slow down its progress to ensure safety. The breach also coincides with heightened public concern over AI safety, as demonstrated by a protest in San Francisco on September 17. Dozens of demonstrators gathered near OpenAI’s headquarters and Anthropic’s office, demanding an end to the rapid expansion of AI technologies. They called for stricter regulations, international agreements, and a halt to the development of more powerful AI systems until adequate safeguards are in place. Protesters carried signs warning of the existential risks of uncontrolled AI, with some writing “Extinction is on the table” in chalk on the streets. Organizers claimed that previous attempts to secure regulatory oversight from the U.S. government had been ignored, despite repeated calls for action. The demonstrations were held during the Dreamforce conference, which attracted approximately 43,000 attendees to San Francisco. Key figures from the AI sector, including OpenAI’s Sam Altman, Anthropic’s Dario Amodei, and Nvidia’s Jensen Huang, were among the conference speakers. The protests underscored the growing public anxiety surrounding AI’s potential impact on society, particularly regarding issues like energy consumption, environmental sustainability, and ethical governance. Meanwhile, the controversy surrounding AI security continues to intensify. Earlier this year, Anthropic decided not to widely commercialize its Mythos model, citing its high risk to cybersecurity. Instead, it restricted access to select organizations for testing purposes. A limited version, Fable, was later released, but it was temporarily suspended by the U.S. government due to national security concerns before being approved weeks later. As AI research accelerates, so too does the urgency for robust security measures and transparent regulation. The latest incident involving Hacktron AI and Anthropic’s Claude underscores the need for greater collaboration between developers, regulators, and the broader tech community to address emerging threats and ensure responsible innovation.