Hacktron researchers used Claude to build an exploit that breached OpenAI’s sign‑in system, gaining access to internal code repositories.
Hacktron researchers leveraged the Claude AI model to automatically generate an exploit that compromised OpenAI’s sign‑in system, ultimately granting the team access to internal code repositories.
How the AI‑Generated Exploit Was Crafted
The researchers fed Claude a series of prompts describing the target authentication flow and asked the model to produce code snippets that could bypass the sign‑in checks. Claude returned a functional payload that, when executed, manipulated token validation and allowed the attackers to masquerade as a legitimate user.
Exploiting the Sign‑In Flaw
OpenAI’s sign‑in endpoint relied on a predictable token structure. By injecting a crafted token generated by the AI, the attackers sidestepped multi‑factor verification and accessed the internal dashboard where source code is stored.
The breach was confirmed when the team observed unauthorized git clone operations from the private repository, exposing proprietary model training scripts and infrastructure configurations.
Implications for AI‑Assisted Security
This incident highlights a growing risk: AI models can be repurposed as offensive tools, automating the discovery and exploitation of vulnerabilities faster than traditional manual methods.
Security teams are urged to treat AI‑generated code with the same scrutiny as any external script, implementing strict validation, sandboxing, and continuous monitoring of authentication mechanisms.
- Review and harden token generation logic
- Implement anomaly detection for unusual sign‑in patterns
- Restrict AI model access to internal documentation
- Conduct regular red‑team exercises using AI tools
AI can be both a defender and an attacker; the line is drawn by how we control its use.
For a detailed account of the research and technical breakdown, see SecurityWeek coverage of AI‑Built Exploit and Sign‑In Flaw.