Researchers used Claude to compromise OpenAI employee accounts
The Verge - AI Surfaced Read the original
According to a Wall Street Journal report cited by The Verge, three independent security researchers at Hacktron used Anthropic’s Claude Opus 4.8 and 5 models to hack into OpenAI employee accounts in under 72 hours. The researchers reportedly gained access to OpenAI’s internal GitHub repository, known as ‘Monorepo,’ which is said to contain proprietary algorithmic material. The account is second-hand, drawn from the Wall Street Journal via The Verge’s summary, and no technical breakdown of the exploited weakness, the specific accounts targeted, or OpenAI’s response has been made public. It is unclear whether this reflects a broadly shared vulnerability in employee account security or a lapse specific to OpenAI’s setup.
rss · The Verge - AI · Sep 18, 15:30
Bug bounty programs assume defenders can outpace AI-assisted attackers
OpenAI, like most major technology companies, relies on authorized bug bounty programs (via platforms such as Bugcrowd) to let external researchers probe production systems and employee-facing infrastructure for weaknesses before malicious actors find them. This model assumes that chaining vulnerabilities into a working compromise still takes attackers meaningful time and effort, giving defenders a window to detect and respond. The Hacktron AI case tests that assumption directly by reportedly using Anthropic’s Claude models to accelerate discovery and exploitation of chained flaws against OpenAI’s own employee accounts and private source code environment, for a reported cost under $3,000.
Who is exposed
This item most directly concerns OpenAI, whose employee accounts and internal source repository were reportedly accessed. More broadly, any organization running large language models with agentic or tool-use capabilities against its own or others’ infrastructure should note that these models were used as an offensive accelerant, meaning the exposure question extends to any company relying on standard employee account protections (passwords, session tokens, single sign-on) without assuming an AI-assisted adversary can compress attack timelines. Organizations should check whether their account compromise detection and credential hygiene practices assume human-paced attackers, since that assumption is what this report puts in question.
What reduces the risk
No technical details of the exploited weakness have been disclosed, so no specific fix can be confirmed; general compensating controls include stronger multi-factor authentication, anomaly detection tuned for rapid automated attack sequences, and tighter access segmentation around sensitive code repositories.
References
Tags: #AI-assisted hacking, #account compromise, #supply chain security, #LLM misuse, #incident disclosure
Commentary on this