Northwestern convenes AI security conference on speed and human accountability
The October 22 gathering examined who can meaningfully challenge AI-driven decisions, alongside research presentations on attack forecasting and safer protein generation.
Listen to this story
The audio brief
Story brief
3 key pointsNorthwestern’s October 22, 2026, national-security AI conference put two questions side by side: whether AI compresses the time and expertise needed to mount threats, and whether human sign-off is meaningful when decision-makers cannot understand or contest system outputs. More than 425 people registered. The program also featured two lab projects, SafePro and NTEWS++, but Northwestern’s recap gives no evaluation metrics; its descriptions of biological safety and six-month attack forecasts are not evidence of validated performance.
- 01
Organizers questioned how accountability and sovereignty work when technology providers own intellectual property, control data, and set standards.
- 02
NSAIL head V.S. Subrahmanian argued that effective early use of AI could give an advantage in defense or malign activity, while warning that threats may differ from predictions.
- 03
The agenda examined whether inexpensive tools such as drones, adaptive bots, open-source penetration tools, and deepfakes let weaker actors impose disproportionate costs.
A human approving an AI-driven security decision may offer little protection if they cannot understand or challenge it. That concern was among the questions examined at Northwestern’s annual Conference on AI & National Security on October 22, 2026. The university’s October 28 recap describes a gathering focused on threats, operational speed and responsibility when decisions go wrong.
Northwestern’s Security & AI Lab, known as NSAIL, convened the day-long event with cohosts Rinn AI and University College Cork. More than 425 people registered to participate in person or online. Discussions covered frontier and military AI in national security, counterterrorism and geopolitical decision-making, bringing technical research into the same program as questions about governance and rules of engagement.
Who can challenge the machine?
One central question concerned nominal oversight: a person approves an AI decision without meaningfully understanding it or pushing back. The conference asked who should be accountable when AI-enabled decisions go wrong, and what rules would allow security operations to move faster while preserving meaningful responsibility. It framed oversight as more than simply having someone available to sign off.
The governance discussion also addressed technology providers’ power. Organizers asked how rules and oversight should work when providers own the intellectual property, control the data and set standards. Their question linked accountability to sovereignty: not only who approves an individual decision, but who controls the technology and information on which that decision depends.
NSAIL head V.S. Subrahmanian emphasized timing, arguing that whoever first uses AI effectively for defense or malign purposes will likely gain an advantage. He described his lab’s work as anticipating adversary behavior and examining defenses, while acknowledging that the eventual threat may differ from the one researchers expect.
My lab predicts what the adversary is going to do before they do it and examines how we might defend against the threat, knowing full well that the threat vector itself is not likely to be exactly what we think it might be,
V.S. Subrahmanian, head of Northwestern’s Security & AI Lab
Cheap threats, costly defenses
Another question was whether AI changes the fundamental nature of security threats or chiefly reduces the time and expertise needed to carry them out. The agenda applied that distinction across war, terrorism, surveillance, malware and influence campaigns. It also asked how defenders should respond when relatively weak actors can impose disproportionate costs using cheap, mass-produced tools.
Examples included drones, adaptive bots, open-source penetration tools and deepfakes. Sessions examined drones and counter-drone policies, social influence operations, technology risk and governance centered on data. The program thus placed the problem of responding to inexpensive threats alongside the separate problem of deciding who has authority over AI-enabled security systems.
Two research presentations, different safety goals
The conference also featured lab research spanning biological design and terrorism forecasting. Northwestern’s recap highlighted two projects, presenting their intended capabilities rather than supplying detailed evaluation results:
- SafePro: NSAIL fellow Marco Postiglione discussed a model for generating proteins. Northwestern describes its output as functional and verifiably non-toxic and non-virulent—claims about both biological usefulness and safety, not merely the ability to generate a protein design.
- NTEWS++: Computer science PhD student Lirika Sola presented the Northwestern Terror Early Warning System. The university describes it as a machine-learning platform that models terrorist behavior to forecast attacks within the next six months. That is the stated prediction horizon, not a reported measure of forecasting accuracy.
Sources
- mccormick.northwestern.eduAnticipating Threats, Ensuring Accountability in AI Security
Reader comments
Newest comments first. Replies stay oldest first.