Security researchers used Claude to exploit OpenAI systems, gaining employee-account and internal-repo access
2026-09-18
TechCrunch reported Sep 18, 2026 that security researchers used Anthropic's Claude to find and exploit vulnerabilities in OpenAI's systems, taking over employee accounts and gaining access to an internal code repository, before reporting the flaws to OpenAI.
Significance 3: A concrete demonstration of a frontier model autonomously discovering and exploiting real vulnerabilities against a major lab's infrastructure is a meaningful agentic-capability and security data point, though it was responsibly disclosed rather than a live incident, holding it at 3.
Implications · machine-drafted, not owner judgment
This is concrete evidence that current frontier agentic models are capable enough to autonomously find and exploit real vulnerabilities against a sophisticated target — useful for AI-assisted red-teaming, but it also lowers the skill floor for real attackers using the same tools. For OpenAI, a researcher-discovered account-takeover and internal-repo exposure is an operational security lapse worth watching for its own public response.
- OpenAI's public acknowledgment or disclosure of the specific vulnerabilities
- Whether similar Claude-enabled exploits are reported against other labs or vendors
Sources
- press Researchers used Anthropic's Claude to hack into OpenAI retrieved 2026-09-18