Sam Altman wants you to know this Pentagon deal is different. Whether you believe him may depend on how much trust you place in a company that has, in recent years, rewritten its own rulebook more than once.
OpenAI's chief executive announced this week that the company has signed a contract with the US Department of Defense, and he was quick to emphasise one thing above all else: the arrangement includes technical safeguards. That framing is deliberate. It is a direct response to the furore that erupted when rival AI company Anthropic inked its own Pentagon deal, prompting fierce internal and external criticism over whether advanced AI systems should be anywhere near military decision-making.
Altman did not spell out exactly what those safeguards entail in technical terms, but the message to critics was clear enough. OpenAI is not handing the keys to an unconstrained system to defence planners. There are guardrails, he says, baked into the architecture of the arrangement itself.
The Anthropic Precedent
To understand why Altman felt compelled to lead with the safeguards pitch, you need to understand what happened to Anthropic. That company, founded by former OpenAI researchers and publicly committed to AI safety, faced a backlash when its Pentagon partnership became public. Critics argued that any arrangement with the defence establishment risked normalising the use of AI in contexts, including lethal autonomous weapons systems, where the stakes of getting it wrong are measured in human lives.
Anthropic pushed back, arguing its contract was limited in scope and that engagement was preferable to leaving the field entirely to less scrupulous actors. It is a utilitarian argument with real force. If advanced AI is going to inform military operations regardless, better that safety-focused companies are at the table than absent from it.
OpenAI is now making a version of the same case, but with Altman apparently hoping the explicit mention of technical protections will pre-empt the loudest objections before they gain traction.
The Bigger Picture for AI Governance
For Australian observers, the significance of this deal extends well beyond American domestic politics. Australia's own defence establishment has been accelerating its engagement with AI technologies, and the frameworks being set in Washington will inevitably shape what becomes acceptable practice among allied nations.
The Australian Parliament has so far been cautious about legislating hard boundaries around AI in defence contexts, preferring to let the technology and its governance frameworks mature before locking in rules. That approach has merit, but it also means Australia's posture on these questions is, in practice, partly outsourced to decisions being made in San Francisco boardrooms and Pentagon briefing rooms.
Here's the thing: the question is not really whether AI will be used in defence applications. That ship has sailed. The genuine debate, the one that matters now, is about accountability, transparency, and where the lines should be drawn around autonomous decision-making in high-stakes environments.
Scepticism Is Warranted, But So Is Nuance
It would be easy to dismiss Altman's safeguards claim as corporate spin, the kind of reassuring language companies deploy when they want to do something commercially lucrative without taking a reputational hit. OpenAI has form here. The company that once embedded a strict non-commercial mission into its founding documents has since restructured in ways that have troubled some of its own former employees and board members.
At the same time, the critics who argue that no AI company should ever engage with defence clients are making an argument that is cleaner in theory than in practice. Governments will build and deploy AI-assisted systems regardless. The relevant question is whether the companies with the most advanced safety research are involved in shaping those systems, or whether the work goes to developers with fewer scruples and less accountability.
There is a legitimate progressive counterargument that participation itself normalises militarised AI in ways that carry long-term risks no technical safeguard can fully address. That concern deserves to be taken seriously, not dismissed as naive idealism.
The honest answer is that both positions contain genuine insight. Blanket prohibition ignores the reality of how defence technology develops. Uncritical participation ignores the very real risk of incremental normalisation. The path between them requires exactly the kind of rigorous, independent oversight that neither the companies nor the governments involved have yet put in place at any convincing scale.
Altman's announcement is a commercial milestone dressed in the language of responsibility. The test of whether that language means anything will come not from press releases, but from independent verification of what those technical safeguards actually prevent, and what they quietly permit.