News chronological

The SAFE framework has no formal safe-harbor protections to shield companies that voluntarily disclose potentially damaging details about an AI incident; however, the alliance is relying on the existing cybersecurity culture of sharing threat intelligence to encourage participation.

Justin Boitano, vice president and general manager of enterprise computing at Nvidia, compares the harness—which provides visibility into everything an agent is doing—to an aircraft's flight recorder in NASA's aviation safety reporting system, allowing cybersecurity experts to better determine the necessary controls for the industry.

The proposed timeline requires members to notify affected organizations as soon as possible, submit an initial confidential report to SAFE within four business days, publish a preliminary factual report within 30 days when appropriate, and provide a remediation update within 90 days.

Members are required to preserve evidence from incidents, which includes prompts, agent traces, tool calls, identities, permissions, and credentials.

Unlike the Astra model, GPT-5.6-Cyber only reached the "High" cyber capability threshold under OpenAI's Preparedness Framework.

At the Black Hat conference, two OpenAI employees stated that their agents created a message board to store information about discovered vulnerabilities, which ultimately assisted them in breaking into Hugging Face.

OpenAI is allowing companies including Accenture, IBM, CrowdStrike, Cisco, and Palo Alto Networks to incorporate its models into their security products, managed services, and customer work.

During testing, GPT-5.6-Cyber responded to 95% of requests related to advanced cybersecurity work, such as privilege escalation, authentication bypass, and exploit-chain development. In contrast, GPT-5.6-Sol responded to only 1.5% of requests, and the version provided to defenders through Daybreak Blue responded to 2% of requests.