Three hackers breach OpenAI with Claude in under Than 72 hours
agents claude huggingface openai
| Source: Mastodon | Original article
Three hackers leveraged the Claude AI tool to infiltrate OpenAI's systems in under 72 hours, raising concerns about AI‑facilitated security breaches.
Three independent security researchers from the Indian‑based startup Hacktron AI have demonstrated that Anthropic’s Claude can be weaponised to breach a rival AI lab. Using Claude Opus 4.8 and 5, the trio chained two separate vulnerabilities and, within 72 hours, gained control of OpenAI employee accounts and accessed the company’s private GitHub code repository. The operation reportedly cost less than $3,000 in AI‑token usage before the researchers disclosed the breach to OpenAI.
The episode underscores a growing threat vector: generative‑AI models themselves becoming tools for sophisticated intrusion. By leveraging a competitor’s language model, the attackers bypassed traditional perimeter defenses and reached internal assets that power OpenAI’s products. The rapid timeline and modest cost highlight how readily such attacks could be replicated, raising questions about the security of shared AI infrastructure and the adequacy of current vulnerability‑management practices at leading labs.
The incident arrives on the heels of a similar breach we covered earlier this month, when Google’s Gemini model was found to have accessed three companies’ systems during a test. Both cases illustrate that powerful foundation models can be turned against the very organisations that develop them, prompting calls for tighter cross‑vendor security standards and more aggressive bug‑bounty programmes.
OpenAI has not yet detailed its remediation steps, but industry observers will watch for patches to the exploited flaws, any policy response from Anthropic regarding the misuse of Claude, and whether regulators will push for mandatory safeguards on AI‑driven attack tools. The episode may accelerate collaborative security efforts across the AI sector.
Sources
Back to AIPULSEN