The Pentagon Wants $30 Million for an AI Lie Detector

NEON DAILY

2026-09-25 · AI · TECHNOLOGY · CULTURE · WORLD

The Pentagon wants $30.3 million over five years to build an AI-assisted lie detector. The program, called Polygraph+ or Polygraph Next in Defense Counterintelligence and Security Agency (DCSA) budget materials, would add machine-learning scoring and "standoff sensing" (physiological readings taken without attaching anything to the subject) to federal polygraph and credibility assessment. Its stated targets are personnel vetting and insider-threat detection, uses where a bad result can carry serious consequences for the person being assessed.

The money is a request, not an appropriation: Congress has not approved it. It also arrives as the Pentagon leans harder on polygraphs. MIT Technology Review, which reported the request, notes that about 50 Joint Staff officers were recently tested as part of an inquiry into press leaks.

DCSA's FY 2027 budget materials describe a modernization effort that includes standoff sensing, automated scoring and decision aids, and centralized data storage and analytics, along with validation studies, prototype evaluations, field testing, and pilots. In other words, the agency is still working toward proof. It is not presenting a finished system with a demonstrated track record in real screening.

That proof is the hard part. A 1983 Office of Technology Assessment review, reproduced by the Federation of American Scientists, concluded that the polygraph instrument does not itself detect deception, and found almost no studies relevant to expanding federal screening. Newer sensors and software don't automatically solve that problem. An algorithm can find patterns in heart rate or skin response, but no one has established a reliable physiological signature of lying for it to find.

The public descriptions of Polygraph Next include no validation results, error rates, demographic-performance data, or operational outcomes. That doesn't mean such data doesn't exist, but it leaves outsiders with no basis for judging whether the system would be accurate, fair, or reliable.

That gap matters because these are not lab exercises. In vetting and insider-threat work, an automated score could shape someone's clearance or career even if officials call it a "decision aid" rather than a final decision. Nothing released so far explains how a score would be weighed, reviewed, or challenged. Until DCSA can answer those questions, $30 million buys a hypothesis, not a lie detector.

Source: MIT Technology Review

ChatGPT and Claude don't just answer questions anymore. Employees are connecting them directly to Notion, Linear, Jira, and the rest of your stack. The AI can read, write, and take actions on company data. Most IT and security teams have no visibility into any of it.

Harmonic Security Connectors changes that. It sits inline with every AI-to-app connection, so you see each call, control what data moves, and block destructive actions before they happen. Employees notice nothing different.

See what's actually running across your business in a live demo.

See what's connected

Owners of some Samsung Bespoke AI 4-Door refrigerators in South Korea say their appliances stopped working, and in some cases stopped cooling, after a SmartThings firmware update on September 22, with dozens of complaints posted to Samsung's Samsung Members community. Samsung says software meant for internal testing was sent to customers by mistake. The company has halted the update and is sending engineers to repair affected units free of charge.

Source: Star News Korea

The Donation of Constantine, a document Lorenzo Valla exposed as a forgery around 1440, helped underpin papal territorial and jurisdictional claims in Italy for centuries. At least 10 popes are known to have cited it in pressing for recognition of papal authority.

Source: Hanover Historical Texts Project

NVIDIA has released Isaac ROS 5.0.0, the latest version of its GPU-accelerated ROS packages, and moved it to ROS 2 Lyrical Luth on Ubuntu 24.04. The update adds ready-made "skills" that let AI coding agents handle robot setup and tuning tasks, and widens deployment across NVIDIA's Jetson hardware. Developers who call NITROS interfaces directly will need to update their code.

Source: NVIDIA

Parallel, Turbopuffer, and Wordsmith run their entire GTM motion on Attio, with agents that chase every buying signal, build pipeline, and move deals forward, 24/7.

Try Attio for free

In a Microsoft Research measurement study, the full software stack for a mobile robot that navigates and handles objects was too heavy to run on smaller onboard GPUs. Larger onboard GPUs could run it but drained the robots' batteries several hours faster. Microsoft has now added tools for offloading that work to edge or cloud GPUs.

Source: Microsoft Research

Scientific American reports that a team of six mathematicians used AI agents to help construct a degree-23 polynomial whose Galois group is M23. It was the last sporadic group not yet shown to arise this way over the rationals, a question open since the 1980s. The researchers credit AI with speeding the search; the final result was checked with conventional math software.

Source: Scientific American

Google's first Project Suncatcher prototype, built with Planet, is scheduled to launch October 1 on SpaceX's Transporter-18 rideshare mission carrying four of Google's TPU AI chips. The mission will test how the chips handle launch stress, radiation, and heat in orbit. Google plans two more prototype satellites in 2027 to test laser links between spacecraft.

Source: Google

Toyota factory workers are training ELEY, a wheeled, two-handed humanoid robot, by performing assembly tasks while wearing grips shaped like the robot's fingers, so the system can capture their movements as training data. The work is part of a much larger plan: starting in 2028, Toyota expects to spend about $6.42 billion a year on factory automation, aiming for roughly 400,000 robots across the group. Toyota says the goal is for robots to work alongside people, not replace them.

Source: Ars Technica

SOURCES

MIT Technology Review · Ars Technica · Microsoft Research · Star News Korea · Scientific American · Google · Hanover Historical Texts Project · NVIDIA

NEON DAILY