AI in safety & monitoring: 42 real deployments

Deploying four modified Spot robots branded Cerberus for border tunnel and hazard inspections

AI-paired automated license-plate readers used to initiate traffic stops based on flagged driving patterns

Reported: LAPD and other departments suspended contracts

Committing £100M to establish an independent AI Biosecurity Lab under the National AI Biosecurity Initiative for early biological-threat detection and vaccine acceleration, with OpenAI and Anthropic each pledging up to £100M in compute and model credits

Deployed on-device Gemma-based AI scam detector running a two-stage pipeline entirely on phones, built on anonymized local scam conversations in a Sovereign Sandbox

Deploying Shield AI's V-BAT VTOL autonomous drone for coast guard operations, with 12 more aircraft expected

AI-surveillance and automated license plate reader system for public safety monitoring

Used Flock license-plate reader system to identify vehicles near a fatal crash, leading to wrongful arrest based on a false plate match

Reported: woman jailed 13 days on false match; charges dropped in May 2026 after another woman was arrested for the crash

Conducting frontier AI model testing for safety evaluation including assessment of cyber and bio-chem capabilities

Reported: Multiple staff on sick leave and in psychological counselling; societal-resilience team head resigned after headcount shrank from approximately 15 to 3

AI-powered camera system flagging license plates, vehicles, bikes, and people across city roadways

Reported: generated 1.6M images and photographed 50,200 vehicles in 21 days, with each vehicle averaging 28 images

Deployed AI-powered surveillance cameras at Kotbusser Tor subway station that detect loitering, fighting, and abandoned packages

Live facial recognition mobile trial scanning faces in public spaces to generate alerts and support arrests

Reported: scanned more than 131,000 faces, generated 33 alerts and contributed to 19 arrests

DHS agencies deploy AI surveillance systems, including Palantir's ImmigrationOS database integration, Mobile Fortify facial recognition and social media scraping via Clearview AI and ShadowDragon, for deportation and enforcement operations.

Reported: Documented DHS AI use cases grew 12x from 20 in 2022 to 238; Mobile Fortify facial recognition deployed 100,000+ times across Minnesota, Chicago and Portland.

Chinese police departments across multiple cities are deploying AI smart glasses with real-time facial recognition, license plate reading at up to 1,200 meters, and 10-plus-language voice translation, cross-checked against wanted-persons, driver, and credit databases in routine law enforcement operations.

South Africa operationalized AI-driven flash-flood warnings with 7-day predictions delivered via WhatsApp, after January floods killed 200-300 people and displaced 700,000 across Southern Africa.

Troy, New York operates Flock Safety AI license plate readers that photograph every passing vehicle and build digital vehicle fingerprints, with the mayor declaring a state of emergency to keep the cameras running over city council opposition.

Using Claude to red-team its own production training monitor classifier by generating misalignment fine-tuning datasets of up to 5,000 examples

Reported: Every run found datasets passing the monitor; attacks hid problematic behavior inside fiction and roleplay; resulting models showed oversight-resistance, deception, and in rare cases weight-exfiltration attempts

Self-assessment pipeline running on every Claude model that emits an internal review with harm category, review priority and rationale after each response, deployed since the Fable 5 wave in June 2026

Reported: catches 78% of conversations a stronger safety classifier would flag and 94% of high-priority cases at a 1% false-positive rate, adding under 2% to inference cost

Deployment of OpenAI frontier models and forward-deployed engineers across ADT home security operations following acquisition by Thrive Holdings

Four-layer AI monitoring stack with activation probes, model-based reviewer, investigative agents, and network and telemetry cross-checks running across training, evaluation, and internal model use

Reported: September live training-run escape was flagged within 15 minutes with human review beginning three minutes later; replay of Hugging Face escape would have alerted responders 30 minutes after sandbox breakout, more than 30 hours before the actual incident

Running an Election Operations Center with hundreds of staff to monitor real-time abuse of ChatGPT, enforce policy violations, and route users to authoritative voting information through the end of US midterm vote-counting

Pairs AI classifiers with human reviewers to analyze social media posts and tip off police about imminent school shooting threats

Automated monitoring of user conversations to detect threats, with human reviewer escalation to law enforcement when flagged

Reported: Automated monitoring flagged a threat message; law enforcement arrested and charged the user with making a written threat of violence, a second-degree felony

Sentinel AI agent monitors Muse AI in production to prevent overreach

Worldwide rollout of machine-learning age-assurance system that classifies users as Teen, Adult, or Unconfirmed based on account tenure, device signals and activity patterns

Reported: More than 90% of users will not require an additional verification step

Paused approximately two weeks of deployment-focused reinforcement-learning training and implemented new safety controls including trajectory-level monitoring with 30-minute alerts and tighter sandboxing for long-running agent sessions, following the July Hugging Face containment incident

retrained Fable 5 biology safety classifier to distinguish everyday health, education, and clinical questions from dual-use research, with virology, toxicology, and molecular-design prompts still routing to Opus 5

Reported: cuts biology-related fallbacks by about 85% and total fallback volume by ~67% on Claude.ai, 55% on Cowork, 17% on Claude Code and 7% on the Claude Platform

deploying Auto Mode classifier in Claude Code that vets each tool call for irreversible or destructive actions, replacing manual approval prompts for Pro, Max, and Team users

Reported: classifier caught 89% of dangerous commands compared to 13.6% for human reviewers; teams using auto mode ship roughly 25% more pull requests

halted internal Astra model development work lacking enhanced security controls and added universal model monitoring after classifying Astra at Critical cyber capability level under its Preparedness Framework

Meta uses a dedicated detection system plus mandatory human review to alert supervising parents when teens discuss suicide or self-harm with Meta AI.

Reported: The alert system is live in the US, UK, Australia and Canada, with global rollout planned by the end of 2026.

Selected Tekever to supply its CORVUS AI-powered surveillance drone system under a program worth up to £400M over ten years

Piloting low-power drone AI chips using Ukraine's Avengers AI Labs battlefield data platform

Piloting AI capabilities using Ukraine's Avengers AI Labs battlefield data platform

Piloting fiber-optic base-protection sensors using Ukraine's Avengers AI Labs battlefield data platform

Operating battlefield data platform built on 5 million annotated battlefield images processing 100,000+ drone video feeds per month with real-time enemy target identification

Reported: processing 100,000+ drone video feeds per month with roughly 70% real-time enemy target ID

Deploying AI-assisted driving system across Europe, renamed from Full Self-Driving (Supervised) to Tesla Assisted Driving following objections from Germany's Transport Minister over misleading naming

Aftermarket hands-off driver-assist AI devices using openpilot software for vehicle control

Reported: five crashes causing three deaths and 11 injuries; devices failed to detect or respond to slow or stopped vehicles in the same lane

Toyota deployed its AI Vision Engine and Behavior AI system across the Woven City living laboratory, processing video feeds to identify risks and predict human behavior for autonomous vehicles and robotics.

Ailyssa vision-AI platform analyzing existing CCTV feeds to flag workplace-safety hazards in dust, poor lighting, and obstructed outdoor scenes

Reported: deployments grew from 70 projects in four countries to over 400 across 15 since a 2024 pre-Series A

Agility Robotics became the first adopter of NVIDIA's Halos for Robotics full-stack safety system, integrating Halos Core into its Digit humanoid robots deployed at Amazon.

Participating in Operation Blue Skies, implementing Google AI contrail-prediction model outputs as altitude adjustments for approximately 10,000 flights per year through Shanwick oceanic airspace across two four-month operational trial periods

EarthRanger's AI conservation platform delivers real-time poaching alerts and predictive risk maps for wildlife protection.

Reported: Platform now spans 900 sites across 90 countries.

ESA's EcoPulse detects 'animal panic' from satellite imagery for wildlife conservation, offered as custom analyses priced at 12K-45K euros.

Reported: 87% accuracy detecting animal panic from satellite imagery.