We Take Over ChatGPT Accounts
claude openai
| Source: HN | Original article
Researchers demonstrated a method to hijack ChatGPT accounts by bypassing its sandbox and content security policy.
A trio of security researchers has demonstrated that OpenAI’s single‑sign‑on (SSO) architecture can be hijacked to seize control of any ChatGPT or Codex account. By chaining two distinct vulnerabilities, the team bypassed the platform’s sandbox and content‑security policy, then used a malicious link to redirect victims through a compromised forum into OpenAI’s authentication flow. Once the SSO path was subverted, the attackers could log in as the target user, gaining full access to the account and any linked services.
The proof‑of‑concept, detailed in a “Do the Math: How We Took Over ChatGPT Accounts” post, showed that the flaw extended beyond ordinary users. The researchers also captured the credentials of several OpenAI employees, opening Outlook, Slack and GitHub sessions and even submitting a harmless pull request that exposed an internal repository. A tweet from the group on July 25 confirmed the breadth of the breach, noting that both employee and unaffiliated accounts were compromised.
Why it matters is twofold. First, the incident highlights a systemic weakness in how OpenAI isolates user identities across its suite of AI tools, raising the risk of data exfiltration, model misuse and supply‑chain attacks. Second, the use of Claude, an external AI assistant, to orchestrate the takeover underscores how inter‑service dependencies can amplify security gaps.
OpenAI has not yet disclosed a formal response, but the episode is likely to accelerate internal audits and prompt patches to the SSO pathway. Industry watchers will monitor whether the company rolls out stricter verification for account‑linking actions, revises its bug‑bounty scope, or faces regulatory scrutiny over the protection of high‑value AI services. The next few weeks should reveal how quickly OpenAI can seal the loophole and restore confidence in its authentication framework.
Sources
Back to AIPULSEN