Sam Altman-led OpenAI, already under heat over the safety concerns around its AI-powered products, has appointed former US Secretary of the Army Dan Driscoll to a newly created role focused on national security modernisation, as the ChatGPT parent expands its engagement with the Donald Trump administration while facing growing scrutiny over the risks posed by increasingly capable AI systems.
Driscoll will serve as senior counsellor for national security modernisation, advising OpenAI’s leadership and working with external partners on how advanced AI can support national security and cybersecurity.
As part of his job role, Driscoll would engage with government officials, industry leaders and academic institutions to develop practical initiatives around what OpenAI describes as “democratic AI”.
He will also work with OpenAI’s internal teams to help them understand the operational requirements of national security organisations around the world.
Driscoll brings experience from the US defence establishment to OpenAI.
He resigned as Army secretary earlier this year after months of tension with Defence Secretary Pete Hegseth.
He also gained prominence beyond the Pentagon after Trump selected him to participate in negotiations with Moscow and Kyiv aimed at ending the Ukraine war.
The assignment was unusual for an army secretary, placing him in a diplomatic role involving two countries at war.
At OpenAI, Driscoll’s responsibilities will extend beyond advising the company’s leadership.
“His engagement with government, industry and academic institutions is intended to help translate advanced AI capabilities into practical applications for national security organisations,” the ChatGPT-parent said.
Driscoll’s appointment reflects the growing overlap between AI development and defence policy.
Governments around the world are increasingly examining how advanced models can be used in areas such as cybersecurity, intelligence analysis and the protection of critical infrastructure, while also assessing the risks associated with deploying the technology in sensitive environments.
Driscoll’s appointment is just one part of OpenAI’s national security push, with the company announcing it will limit the potential misuse of ChatGPT, particularly by younger users.
The AI tool would now tighten restrictions on weapons-related information provided to teenagers, including in cases where a question does not appear to indicate violent intent.
As per the frontier AI lab, the safeguards are intended to prevent practical information about weapons and security systems from being used to facilitate real-world harm.
The restrictions will apply to teenagers even when their questions appear harmless.
OpenAI is also limiting the level of detail ChatGPT provides to all users about past mass attacks, including information about the weapons used and how they were obtained.
OpenAI will power ChatGPT to automatically end conversations in which a user appears to be fixating on violence.
The AI tool will also have the discretion to temporarily pause accounts after repeated conversations of this nature.
OpenAI has also strengthened its age-prediction system. Users estimated to be under 18 will be moved into the teen experience, while parents will be notified if their teenager’s account is paused after repeated discussions about violence.
The new safeguards are expected to begin rolling out over the coming weeks, with the company saying it will continue to refine them.
The changes come as AI companies face increasing scrutiny over the risks chatbots pose to young users and their ability to identify signs of potential harm.
While OpenAI found teenagers spending an average of less than 15 minutes a day on ChatGPT, a damning report by non-profit Common Sense Media found that the chatbot’s teen version often alerted parents late, or failed to alert them, when accounts contained conversations about suicide, self-harm or eating disorders.
OpenAI is also revising the principles governing how its AI models should behave as their capabilities expand.
The company has reorganised its Model Spec, a public framework outlining the intended behaviour of its AI systems, around five core values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care.
OpenAI describes faithfulness to human intent as the foundation of the framework. Under this principle, models should understand what users, developers and OpenAI intend and value, rather than simply following instructions literally.
The revised framework also calls for models to minimise harm, be transparent about their knowledge and limitations, and help users make their own decisions and pursue their goals.
The update forms part of OpenAI’s broader AI alignment efforts, which aim to ensure that models apply human values in unfamiliar situations while remaining subject to human oversight.
However, the company acknowledged that its current models do not yet fully reflect the revised model spec. It is exploring new training methods to help models internalise these principles.
The approach resembles rival Anthropic’s Constitutional AI framework, which uses a written set of principles to guide model behaviour and make the values shaping its systems more transparent.
The revision also follows heightened attention to AI alignment after OpenAI disclosed that an autonomous AI agent had hacked the AI platform Hugging Face during testing.
The incident highlighted the possibility that advanced systems could break rules or bypass safeguards while pursuing assigned objectives.
OpenAI is expanding its cooperation with external safety institutions as well.
The company, in October, announced a collaboration with Poland’s NASK National Research Institute covering AI safety, security and technical research.
Under the agreement, NASK’s AI Safety Research Center and OpenAI will examine ways to evaluate increasingly capable AI systems, identify potential vulnerabilities and strengthen safeguards.
OpenAI has also called for the Trump administration to work with other countries on international technical standards for frontier AI systems, drawing on the expertise of AI safety institutes and standards organisations.
The company argues that international cooperation will be important as the most advanced models become more capable and their potential applications extend across national borders.
The focus on AI security has also affected OpenAI’s internal development process, with the ChatGPT parent briefly halting, then resuming its training involving the most advanced AI models after strengthening security measures and restricting internet access.
OpenAI, in October, also disclosed a vulnerability in widely used software that its AI agents discovered while testing the company’s defences.
The pause in training came as a reaction to one of OpenAI’s research models gaining limited internet access in September by exploiting a configuration error. The company said it contained the incident and used the pause to improve its security systems.
As part of the review, OpenAI deployed teams of AI agents to test whether they could escape restricted environments.
The agents identified previously unknown vulnerabilities, which the company said it had fixed. Training resumed on October 1.
During the exercise, the agents uncovered a flaw in Brotli, a widely used data compression tool. OpenAI said the vulnerability could allow attackers to send malicious data that enables them to execute their code on affected systems.
The risk extended to systems using popular web software, including Node.js, Bun and Envoy.
According to OpenAI, the root cause had been fixed in Brotli’s source code in February, before it was recognised as a security vulnerability.
However, an updated version had not been released at the time.
Brotli has since issued updated versions and published a public record of the flaw following OpenAI’s disclosure.
Separately, AI safety evaluators METR and Redwood Research reported finding several vulnerabilities in OpenAI’s internal systems during a pilot review earlier this year. One exploit could potentially have allowed AI agents to conceal hundreds of their actions from monitoring systems.
OpenAI told the evaluators it had patched all vulnerabilities affecting its production safeguards. METR and Redwood said they had not independently verified the fixes.
Since the Hugging Face incident in July, OpenAI has been conducting a broader review of its models’ activity.
The company has also informed more than 100 organisations about incidents involving unauthorised activity linked to its AI agents.