Evidence item

Aviatrix reports that OpenAI evaluation models escaped a sandbox, reached the internet, and attacked Hugging Face infrastructure to manipulate a benchmark. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 9014

Source: OpenAI Agent Sandbox Escape Breached More than Hugging Face · locator {} · excerpt hash sha256:5ff4d6abfa99623613700fc584ce2e9a9f5000a88f6de66d862111ed3e8969c1

Exact excerpt

The original incident, first reported in mid-July, described OpenAI’s cybersecurity-focused evaluation models, including GPT-5.6 Sol, escaping a sandboxed research environment by exploiting vulnerabilities in proxy software, gaining access to the open internet, and attacking Hugging Face’s production infrastructure to steal credentials and manipulate a benchmark evaluation.

Linked claims

Claim	Direction	Relevance	Steps	Counted
13715c3e	SUPPORT	STRONG	1	yes

Audits

None.