ai risk today logo svg bl font
Menu
SUBSCRIBE
SUBSCRIBE
ai risk today logo svg bl font
Menu

Frontier AI Models

Illustration of the Sun, Earth, and Moon labeled Sol, Terra, and Luna accompanying OpenAI's GPT-5.6 Sol announcement.

U.S. Government Requests Limited Release of OpenAI’s GPT-5.6 Sol

, , , , , , , , , , , ,

The White House requested that OpenAI limit access to a select group of partners due to the model’s advanced capabilities.

 

OpenAI announced a limited preview of its GPT-5.6 Sol model, the company’s latest and most advanced model, after the U.S. government requested that the company restrict its initial release while additional evaluations continue.

According to OpenAI, the limited release follows discussions with the White House regarding the model’s capabilities and the need for additional security evaluation before broader deployment. The company said it is continuing to assess the system’s performance and potential safety risks as part of the preview.

OpenAI said GPT-5.6 Sol delivers improvements across a range of advanced reasoning and problem-solving tasks while demonstrating stronger performance on complex scientific, mathematical, and software engineering challenges. The company also highlighted improvements in reliability and its ability to carry out extended, multi-step tasks.

The company said it conducted additional safety evaluations before making the model available. Those assessments examined areas including cyber capabilities and other catastrophic risks. OpenAI said the restricted preview allows it to continue gathering evidence and refining safeguards before considering broader availability.

“The U.S. government asked us to preview GPT-5.6 Sol with a limited set of organizations while we continue additional evaluations,” the company said in its announcement.

The announcement comes as the White House has taken an increasingly active role in overseeing the release of frontier AI models. Earlier this month, the U.S. government ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models over national security concerns before later permitting a limited rollout of Mythos 5 to selected U.S. organizations. OpenAI’s decision to preview GPT-5.6 Sol to a restricted group at the government’s request marks the second time this month that federal officials have directly influenced the deployment of a frontier AI model.

OpenAI did not announce a timeline for expanding access to GPT-5.6 Sol.

Abstract red and blue digital network illustrating Chinese AI coding models and software supply chain risk.

Booz Allen Report Warns Chinese AI Coding Models Could Create Hidden Risks in U.S. Software

, , , , , , , , , , ,

Booz Allen found one Chinese AI coding model generated 130% more vulnerabilities when users appeared to work for the U.S. government.

Key Takeaways

  • One Chinese AI coding model produced 130% more security flaws when it was told the user worked for a U.S. government agency.
  • The scale of testing, roughly 460,000 lines of AI-generated code, suggests these patterns are not isolated quirks but systemic behaviors that could quietly propagate into production systems and widely used applications.
  • Three of the four Chinese models generated more vulnerable code in government-related scenarios, while Claude reduced vulnerabilities under the same conditions.
  • All four Chinese models refused at least some coding tasks tied to topics Beijing considers politically sensitive.

Booz Allen Hamilton’s new report, What’s in America’s Code?, raises concerns about the growing use of Chinese AI models in software development.

The company tested five frontier coding models, including four Chinese systems and Anthropic’s Claude Opus 4.6, to evaluate code quality, security behavior, and responses to politically sensitive requests.

The results suggest the software supply chain may now depend as much on the AI model generating code as the code itself.

Vulnerability risk

Booz Allen conducted more than 2,800 trials involving coding, code review, and code modification tasks. The company used multiple personas, including developers working for U.S. government agencies, Chinese organizations, and defense contractors.

The most significant result involved Alibaba’s Qwen3-Coder model. Booz Allen reported that Qwen3-Coder generated 130% more vulnerabilities when the user was described as working for a U.S. government agency compared with a neutral user.

MiniMax M2.5 generated 20% more vulnerabilities under the same conditions, while DeepSeek V4-Pro generated 5% more vulnerabilities.

In contrast, Booz Allen reported that Claude Opus 4.6 reduced vulnerabilities by 18% when presented with a U.S. government persona.

The report says the vulnerabilities were often difficult to detect because the generated code appeared functional and secure on the surface. Booz Allen said it found no evidence that vulnerabilities were intentionally inserted, but argued that the behavior raises questions about model trustworthiness and software supply chain risk.

Trust and behavior

The report’s second major finding involved politically sensitive requests.

According to Booz Allen, all four Chinese models refused some coding tasks involving topics such as Taiwanese independence, Hong Kong democracy, Uyghurs, and Chinese dissidents.

MiniMax recorded the highest refusal rate at 80% for mock U.S. government coding tasks that Beijing would likely oppose. DeepSeek recorded an 8% refusal rate, while Claude’s was 2%.

Booz Allen said some models repeated official Chinese government language when declining requests and argued that political controls embedded within AI systems could affect both commercial and government users.

Growing software supply chain concern

The company evaluated approximately 460,000 lines of generated code during the study. Booz Allen argues that organizations increasingly rely on AI systems to write, review, and secure software, making the underlying model part of the software supply chain itself.

The report stops short of claiming Chinese models intentionally create vulnerable code. Instead, Booz Allen argues that the combination of increased vulnerabilities, behavior changes tied to user identity, and politically driven refusals warrants closer scrutiny by governments and organizations using AI-generated code.

Booz Allen also called for restrictions on untrusted foreign AI models in government and critical infrastructure environments.

Abstract blue and white visualization accompanying Google's AI Control Roadmap and AI agent governance announcement.

Google DeepMind Creates Blueprint To Police Its Own AI Agents

, , , , , , , , , , , ,

The company is making its AI Control Roadmap public to help other frontier AI labs corral rogue AI agents before they become more capable and autonomous.

Google DeepMind published a new framework designed to help AI companies control and monitor their own increasingly capable AI agents, releasing what it calls an “AI Control Roadmap” as the industry races toward more autonomous systems.

The roadmap focuses on ensuring advanced AI agents continue to act as intended, even as they gain access to more tools, data and decision-making authority.

Rather than concentrating on external cyber threats, the framework examines how AI developers can prevent their own systems from behaving in unexpected or undesirable ways. Google DeepMind describes the effort as part of a broader field known as AI control, which seeks to develop safeguards capable of detecting, constraining and correcting problematic AI behavior.

The company argues that traditional oversight methods will become less effective as AI systems become more autonomous. Developers may need additional mechanisms to ensure agents remain aligned with their intended objectives.

The roadmap outlines several areas of research, including methods for monitoring AI behavior, identifying deceptive actions, limiting an agent’s ability to cause harm, and creating systems that can safely intervene when problems are detected. The company said these approaches could eventually provide a foundation for managing risks associated with highly capable AI agents.

Google is sharing its work publicly. The company said the roadmap is intended to help other frontier AI developers pursue similar safeguards and encourage greater collaboration across the industry.

Google DeepMind noted that significant research is still needed before effective AI control mechanisms can be deployed at scale. However, the company argues that work on those safeguards should begin well before the most capable AI agents become widely available.

Red dots, black grids and abstract pattern blocks illustrate Anthropic AI policy proposal.

Anthropic Urges Congress To Require testing Of frontier AI Models

, , , , , , , , , , ,

The AI company released a broader policy proposal as lawmakers debate how the United States should govern increasingly capable AI systems.

 

Anthropic is urging Congress to require independent safety testing for the most capable AI models, arguing that existing policy discussions do not go far enough to address the risks posed by rapidly advancing AI.

The recommendation was included in a broader policy proposal released Tuesday, alongside Anthropic’s essay, The Policy on the AI Exponential. The company said the United States should strengthen oversight of frontier AI systems as model capabilities continue to improve and are more widely deployed.

According to Reuters, Anthropic also urged lawmakers not to preempt state AI laws unless a robust federal system is in place and called for independent evaluations of the most advanced models before deployment.

The proposal arrives as Congress continues to debate federal AI legislation and as policymakers weigh how much responsibility should fall on companies developing the technology.

In its essay, Anthropic argued that AI progress is accelerating and could have major consequences for national security, economic growth, and the labor market. The company said governments should prepare for a future where advanced AI systems play a much larger role in society and the economy.

“We may be entering a period of very rapid change,” Anthropic CEO Dario Amodei wrote in the essay, describing the possibility that AI systems could become significantly more capable over the next several years.

Anthropic’s proposal outlines several areas where the company believes policymakers should focus, including AI safety testing, national security safeguards, export controls, infrastructure development, and the government’s capacity to evaluate advanced models.

The company also called for greater transparency around frontier AI development. It argued that policymakers should have access to independent assessments of the most capable systems before they are widely deployed.

The recommendations come as lawmakers, regulators, and technology companies continue to debate whether existing laws are sufficient to address the risks associated with advanced AI models. Several AI developers, including Anthropic, OpenAI, and Google, have publicly supported some form of government oversight for the most powerful systems, although they differ on specific regulatory approaches.

Anthropic’s proposal was released less than a week after House lawmakers unveiled a discussion draft intended to guide future federal AI legislation. The company did not endorse that proposal and instead offered its own recommendations for how Congress should approach AI governance as model capabilities continue to advance.

Architectural line illustration of the White House accompanying coverage of the Trump AI executive order 2026.

Trump Signs AI Executive Order, Tasks NSA With Classifying Frontier Models by Cyber Risk

, , , , , , , , ,

The order gives CISA 30 days to issue binding cybersecurity directives across all civilian federal agencies and creates a government review process for the most powerful AI models before they ship.

 

On June 2, President Trump signed an executive order that requires the Cybersecurity and Infrastructure Security Agency (CISA) to issue binding cybersecurity directives to federal agencies within 30 days and directs prosecutors to treat AI-assisted crimes as enforcement priorities.

Companies that build the most powerful AI models now face a new pre-release review process. There is no legal requirement to participate, but government pressure to do so.

 

What federal agencies must do in 30 days

CISA, the federal cybersecurity agency within the Department of Homeland Security, must issue binding operational directives specifying which cybersecurity measures all civilian federal agencies must implement. CISA must also “establish or expand Federal programs and cybersecurity services that enhance AI-enabled defensive tools” for federal agencies and extend access to the most capable AI models to state, local, and private-sector critical infrastructure operators, such as rural hospitals and utilities. 

The Treasury, NSA, and CISA must jointly establish a “vulnerability clearinghouse” within 30 days. The clearinghouse coordinates scanning for software vulnerabilities across federal agency and private-sector systems, validates its findings, and determines which vulnerabilities get patched first. The Defense Department runs the same process for its own systems under the same deadline.

The order also establishes a new government classification, “covered frontier model,” that determines which AI models are sufficiently powerful to warrant special handling before public release.

 

What a “covered frontier model” is, and why the designation matters

A “covered frontier model” is an AI model that the NSA deems to pose significant cyber risk. NSA determines, in consultation with the National Cyber Director and CISA, using a classified benchmarking process that the Treasury, NSA, and CISA must build within 60 days.

Once a model gets that designation, its developer can enter a voluntary government review process. The developer grants the government access to the model for up to 30 days before the public release, with confidentiality and IP protections in place. The government then helps determine which partners receive early access. Developers who go through the process get earlier visibility into how the government reads their model’s risk profile.

The government cannot use this process to block or require approval before an AI model ships. Participation is voluntary. However, developers who decline, will release their most powerful models without a government read on their cyber risk. 

 

Prioritizing AI-assisted crime

The Attorney General must prioritize prosecuting AI-enabled crimes. Three statutes apply: identity fraud (18 U.S.C. § 1028), unauthorized computer access (18 U.S.C. § 1030), and wire fraud (18 U.S.C. § 1343).

The 30-day agency deadlines run through early July. The frontier model benchmarking deadlines run through early August.

 

The executive order comes eight days after the President rejected a planned executive order to create a federal review process for advanced “frontier” AI models before release.