OpenAI grants Washington early access to GPT-5.5 model

Chris Lehane, Executive Director of Global Affairs at OpenAI, announced that the company has granted U.S. authorities early access to its latest artificial intelligence model, GPT-5.5, as part of an updated agreement with the American Institute for AI Standards and Innovation (CAISI).

Lehane wrote on the LinkedIn platform: "As part of an updated agreement with the American Institute for AI Standards and Innovation (CAISI), we provided early access to OpenAI's advanced systems, including GPT-5.5, for testing and evaluation in the interest of national security." He explained that the company is also cooperating with U.S. authorities in testing its future models, such as GPT-5.5 Cyber, a specialized model designed to enhance capabilities in the field of cybersecurity.

Similar agreements with major companies

OpenAI was not alone in this approach. The U.S. Department of Commerce announced on Tuesday that Microsoft, Google DeepMind, and xAI (affiliated with Elon Musk) signed similar agreements to share their advanced models with the U.S. government for national security reviews.

Chris Fall, Director of the American Institute for AI Standards and Innovation (CAISI), said in a statement: "Rigorous, independent standard science is essential for understanding advanced AI and its implications for national security. These expanded industry partnerships help us scale our work to serve the public interest at a critical moment."

Shift in Trump administration policy

These developments came a day after a report published by The New York Times (Monday), which revealed that the Trump administration is considering imposing government oversight on new AI models before they are made available to the public, a sharp shift from the previous hands-off approach. Anthropic had raised concern in government circles after launching the "Mythos" model, a model with advanced capabilities in identifying security vulnerabilities that could be used, if it falls into the wrong hands, to launch devastating cyberattacks.

This move represents a major shift in the Trump administration's policy, which had pledged at the beginning of its term to liberate AI from government restrictions. This change comes amid increasing pressure from both the Republican and Democratic parties in Congress, and growing concerns about the impact of AI on national security and jobs.

Assessments of GPT-5.5 capabilities

Assessments conducted by the British AI Safety Institute (AISI) revealed that GPT-5.5 demonstrated advanced capabilities in executing cyberattacks in a controlled research environment. The model showed an ability to solve high-difficulty tasks with a success rate of 71.4%, surpassing the "Mythos" model which had a success rate of 68.6%. The report stated that in a simulation of a real network environment comprising 32 stages, the model managed to successfully complete a penetration task in two out of 10 attempts.