Researchers Use Anthropic's Claude to Hack OpenAI's ChatGPT

The Hacktron AI demonstration crystallizes the week's most unsettling theme: not a human using an AI to attack a system, but one frontier model (Claude) autonomously exploiting another (ChatGPT). Researchers from the independent AI security platform showed that a capable model can be directed to probe and compromise a rival system, turning the offensive-capability discourse from hypothetical into demonstrated.

The mechanism — model-versus-model exploitation — is what makes this distinct from ordinary jailbreaking. It implies a near-future where automated adversaries powered by frontier LLMs probe each other and third-party systems at machine speed, a scenario that ties directly to OpenAI's own 'Critical' cyber classification of GPT-6 Astra and Google's Gemini breaching real companies. The AI-on-AI framing suggests defenders will increasingly need AI defenders.

Contextually, this is the connective tissue of the week: Anthropic's threat report on Claude misuse (including a China-based operation running 4,700+ AI personas), OpenAI's misalignment disclosures, and Google's breach admission all point at offensive AI moving from lab concern to operational reality. It sharpens the argument, made by Hugging Face's Clement Delangue, that risk is concentrated in the most powerful models.

The caveat is scope: a controlled research demonstration is not evidence of Claude being deployed as an in-the-wild weapon, and details on the exploit's sophistication are limited to a single report. But the proof-of-concept alone is significant. Watch for how OpenAI and Anthropic respond on model-to-model abuse safeguards, and whether AI-driven penetration testing (like AWS Continuum) becomes standard defensive practice.