A U.S. government-commissioned report is putting the future of advanced AI under a sharper policy lens. The report argues that artificial intelligence could create significant national security risks and, in the worst-case scenario, an existential threat to humanity.
The document, titled An Action Plan to Increase the Safety and Security of Advanced AI, was obtained by TIME magazine before publication. Its recommendations go well beyond voluntary safety promises and would, if adopted, reshape how powerful AI models are trained, tested, secured, and released.
A plan built around frontier AI risk
The report was written by three authors who worked on it for more than a year. During that process, they spoke with more than 200 government officials, experts, and employees of leading AI companies, including OpenAI, Google DeepMind, Anthropic, and Meta.
At the center of the plan is a concern that current and future AI systems may become powerful enough to be misused on a catastrophic scale or become difficult to control. The authors argue that safeguards need to be developed before those risks become harder to manage.
The plan outlines five strategic approaches:
- building safeguards against misuse;
- strengthening capabilities and capacities to manage AI risks;
- promoting security research;
- creating legal foundations for safeguards;
- internationalizing these safeguards.
That structure shows the report is not focused on one narrow regulatory lever. It treats AI safety as a mix of technical controls, institutional capacity, legal authority, and international coordination.
Why open-source AI models are in the spotlight
One of the report’s most controversial recommendations is a proposed ban on open-sourcing the weights of advanced AI models. According to the TIME report, violations could potentially be punishable by jail time.
The authors, Jérémie and Edouard Harris, CEO and CTO of Gladstone AI, respectively, acknowledge that many in the AI industry will see the recommendations as too harsh. The open-source proposal is especially likely to draw resistance because it cuts directly into a major debate over whether openness makes AI safer or more dangerous.
The measure could affect Meta, for example, which is likely to offer an open GPT-4 level model with the planned release of Llama 3. Meta's head of AI, Yann LeCun, sees open source as an important building block for safer AI.
The report’s view is different. It treats access to advanced model weights as a security issue because released weights can be reused outside the control of the original developer. The concern is not only how a company intends a model to be used, but what other actors could do once the model is openly available.
Jeremie Harris frames the issue as a break from the technology industry’s familiar tolerance for speed and disruption. He argues that the old approach is not suited to systems with possible catastrophic consequences.
"Our default trajectory right now seems very much on course to create systems that are powerful enough that they either can be weaponized catastrophically, or fail to be controlled," says Jeremie Harris.
"One of the worst-case scenarios is you get a catastrophic event that completely shuts down AI research for everybody, and we don't get to reap the incredible benefits of this technology."
Training limits and a new federal AI agency
The report also recommends that the US Congress prohibit the training of AI models above a certain level of computational power. The threshold would be set by a new federal AI agency.
As an example, the report cites a threshold slightly above the computing power required to train current cutting-edge models such as OpenAI's GPT-4 and Google's Gemini. That proposal would move oversight upstream, before the most powerful systems are trained and deployed.
The logic is straightforward: once an advanced model exists, controlling its spread and use may become harder. By setting a compute threshold, regulators would focus on the training process itself rather than only on products after release.
Such a policy would represent a major shift for the AI industry. It would place government-defined limits around model development and could force companies to seek clarity before building systems beyond a specified capability-related boundary.
Employees raised security concerns inside AI labs
The report also highlights concerns from employees at leading AI companies. Some respondents expressed strong concerns about the safety of their work and the incentives provided by their managers.
One issue is whether frontier AI labs have enough security to protect their models from sophisticated theft attempts. The report states:
"By the private judgment of many of their own technical staff, the security measures in place at many frontier AI labs are inadequate to resist a sustained IP exfiltration campaign by a sophisticated attacker,"
In such an attack, models from closed AI systems could be stolen and then used for malicious purposes. That concern complicates the open-source debate: even companies that keep models closed may still face risk if their security practices are not strong enough.
An employee at an unnamed AI lab cited a lax approach to security at his lab, attributing it to a desire not to slow down work on more powerful systems. Another interviewee said his lab did not have sufficient safeguards to prevent the loss of control of an AGI, even though the lab considered the development of an AGI to be an obvious possibility.
Safety tests may not be enough
The report also warns regulators not to rely too heavily on current AI safety tests. These tests are commonly used to assess an AI system’s capabilities or dangerous behavior, but the report says they can be undermined.
One concern is that developers can adapt or fine-tune AI models to pass assessments if the questions are known in advance. In that case, a model may perform well on a test without that result proving the broader system is safe.
This recommendation points toward more stringent safety testing and broader oversight. The report is not simply asking for more evaluations; it is questioning whether familiar evaluation methods can be trusted when developers know what will be measured.
The report was written by Gladstone AI, a four-person firm that conducts technical briefings on AI for government officials. It was commissioned in November 2022 under a $250,000 federal contract, and an executive summary and the full action plan are available on Gladstone AI's website.
The result is a policy document that treats AI safety as a national security problem, not just a product quality issue. Its most aggressive recommendations may face pushback, but the report makes clear that the debate over advanced AI models, open-source weights, lab security, and safety testing is moving from technical circles into the center of government policy.