Researchers used Anthropic’s Claude to hack into OpenAI

Security researchers demonstrated that Anthropic's Claude model can be used offensively to discover and exploit security vulnerabilities. The team used the AI assistant to take over OpenAI employee accounts and gain access to an internal code repository.

This incident highlights a growing concern in AI safety: large language models can assist in sophisticated cyberattacks, not just benign coding tasks. The ability of an LLM to chain together exploit steps and navigate real-world systems shows the dual-use risk of increasingly capable models.

The researchers responsibly disclosed the vulnerabilities to OpenAI, meaning the flaws can be patched. However, the exercise proves that AI tools from one lab can be turned against the infrastructure of another, raising questions about how AI providers should secure their own systems against AI-assisted attacks.

For the broader AI industry, this underscores the urgency of red-teaming and access controls. As models become more autonomous and capable of multi-step reasoning, providers will need to assume that adversaries can use rival AI systems to probe their defenses.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.