Since the Pentagon's own IT chief flagged AI-driven cybersecurity vulnerabilities at the TechNet Cyber conference last week, the legislative battle over military AI has accelerated sharply — and the two sides could not be further apart.

The Bills on the Table

Sen. Kirsten Gillibrand, D-N.Y., announced the Secure and Accountable Military AI Act on June 2, according to Defense One. The bill would impose hard restrictions on fully autonomous weapons, domestic surveillance, nuclear targeting, and cyber operations — designating them "high consequence" uses requiring written sign-off from an undersecretary or the Joint Chiefs vice chairman.

Gillibrand is pushing elements as amendments to the Senate's National Defense Authorization Act, which the Senate Armed Services Committee is scheduled to mark up next week, she confirmed to Defense One.

Sen. Elise Slotkin, D-Mich., is doing the same with a separate AI guardrails bill she introduced earlier this year, according to NOTUS as cited by Defense One. Sen. Mark Kelly, D-Ariz., told the Brookings Institution back in March — per Nextgov/FCW — that he's been working with colleagues on NDAA language governing AI use in military operations since the Anthropic-Pentagon dispute.

So at least three Senate Democrats are trying to legislate military AI constraints simultaneously. None of these bills has Republican co-sponsors on record.

Trump's Counter-Move

The same day Gillibrand introduced her bill, President Trump signed an executive order directing that AI be "deployed rapidly to confront any and all threats" to the United States, according to Defense One. The order explicitly rejected what it called "overly burdensome regulation" as a threat to American AI dominance.

The conflict is direct. The White House wants speed. Senate Democrats want guardrails. The NDAA is where those two forces are about to meet.

What's Actually at Stake

The Brennan Center for Justice laid out what the current NDAA already does — and it's thin. The existing law requires intelligence agency heads to track AI performance on criteria like "efficacy, safety, fairness, transparency, accountability, appropriateness, lawfulness, and trustworthiness" under Section 6602. Section 6603 requires testing standards for commercial AI tools like ChatGPT or Google Gemini used by agencies.

The Brennan Center calls this framework "threadbare" compared to guidance the Biden White House issued in 2024. Vague tracking requirements without enforcement teeth are bureaucratic theater.

Meanwhile, the cybersecurity threat driving this whole debate is real and documented. David Markowitz, deputy chief information officer for the Army, told the TechNet Cyber conference that AI tools are "lowering the barrier of entry and exposing more of our attack surface" on the Army's unified network. The Army spent years consolidating its previously siloed IT architecture into a single unified network — and now that unified structure means a single breach point matters more than ever.

Markowitz's assessment, reported by Breaking Defense, was blunt: the Army needs to "ingest the threat and act faster than any adversary." He also said the hardest barrier to fixing this is cultural — bureaucratic compliance mindsets rather than operational ones.

Leidos executives Paul Welch and Josh Salmanson told Breaking Defense the problem starts at system design. Too many military systems are built fragile because mission demands never allow downtime for security improvements. When patches get rushed without proper testing, they can break other critical functions. This is the current operating reality.

The Partisan Frame Misses the Point

Most mainstream coverage frames this as a simple partisan fight: Democrats want guardrails, Republicans want speed.

The real tension is operational vs. political. Sen. Kelly himself told Brookings that combat will sometimes require autonomous offensive systems operating WITHOUT a human in the loop — and he's a Democrat and a former astronaut and Navy combat pilot. He said exceptions will have to be made. That's a more honest position than Gillibrand's bill implies.

Conversely, Trump's executive order demanding rapid AI deployment doesn't address the Army's own admission that its networks are already being outpaced by AI-enabled attackers. Deploying more AI faster into a network with known vulnerabilities isn't bold — it's reckless.

The Anthropic dispute hanging over all of this — where the AI company objected to how the Pentagon intended to use its technology — has never been fully resolved in public. That's a story mainstream outlets keep referencing without explaining. Anthropic builds the Claude AI model and reportedly pushed back on certain military applications. The full terms of that dispute remain opaque.

What Comes Next

Regular Americans don't control autonomous drones or AI targeting systems. But they fund them — and they live in the country those systems are supposed to defend. If AI makes it easier for adversaries to penetrate military networks, that's a national security problem. If AI autonomously makes lethal decisions without human accountability, that's a legal and moral problem.

The NDAA markup next week will be the first real test of whether Congress has the spine to draw any lines at all — or whether they hand the Pentagon a blank check and call it strength.