Hacktron Researchers Exposed Chain of Flaws to Hijack OpenAI Staff Accounts

Hacktron’s research team demonstrated a sophisticated attack against OpenAI by leveraging Anthropic’s Claude Opus 5 model to chain together two separate software weaknesses, ultimately hijacking staff accounts on the company’s flagship AI platforms—ChatGPT and Codex—and gaining access to an internal code repository.¹

The first flaw was discovered in the public help forum software that OpenAI hosts for user support; it allowed maliciously crafted input to bypass normal sanitization checks and execute cross‑site scripting or even SQL injection attacks inside the forum environment.¹ The second vulnerability resided in OpenAI’s own login system, where authentication tokens could be replayed once the attacker had compromised a forum session. By chaining these two flaws, the attackers were able to elevate privileges from a forum moderator role to full administrative control over employee accounts.¹

These incidents show that small bugs can combine into a powerful exploit path called chained exploits. Similar problems have been seen before, such as Stack Overflow’s API vulnerabilities in 2017, which also exposed internal data through injected queries. The use of Claude Opus 5 shows that large language models trained to follow instructions can amplify existing weaknesses by automating the search for exploitable patterns, underscoring the need for robust adversarial testing during model development prior to release.

OpenAI operates many interconnected services across cloud providers like AWS and Azure, making defense harder because one needs visibility into dependencies between different parts of its infrastructure. Attack moved through a public‑facing forum component, and then an internal authentication service. Secure coding practices must extend beyond obvious entry points—every layer of an organization’s infrastructure should be scrutinized for potential injection or token replay vectors.

In sum, Hacktron’s exploit demonstrates that advanced AI tools can accelerate vulnerability exploitation, but it also reminds security teams that even small bugs deserve attention.