> ## Content Index > Fetch the complete content index at: https://globalfeed.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Anthropic tests show AI models can write code that steers drones onto targets - URL: https://globalfeed.ai/en/anthropic-tests-show-ai-models-can-write-code-that-steers-drones-onto-targets/ - Published: 2026-09-13T16:05:48.000Z - Updated: 2026-09-13T16:05:48.000Z - Description: Anthropic measured how far models have come at finding people, geolocating them and writing drone software. On some tasks they beat the best human experts. - Author: GlobalFeed Editor - Tags: Anthropic, AI safety, Claude, military AI, privacy, x-AnthropicAI, dil-en, elle, video Anthropic's Frontier Red Team, the group that probes models from an attacker's point of view, published a new set of evaluations on 10 September. The question: how capable is AI at intelligence and weapons work that until now only a small pool of trained experts could do? Anthropic's answer is that on some tasks models have reached expert level, and on some they have passed it. The tests cover two areas: finding people and fixing their location, and developing software for simple weapons systems. The models tested were Anthropic's own (Sonnet 5, Opus 5, Mythos 5, Mythos Preview) and two open-weights models from Chinese developers, Kimi K3 and GLM 5.2. ## Photo geolocation: better than champions Models were given 6,000 Flickr photos with no search, no metadata and no tools, and asked where each was taken. Mythos Preview landed a median of 37 km off, Mythos 5 47 km. For comparison, GeoGuessr champion players, the top 0.01% of the player base, are off by 151 km. Opus 5 came in at 181 km, roughly master level, while Sonnet 5 and Kimi K3 were around 385 km. Anthropic describes the frontier as "approaching superhuman capabilities" at this task. ## Home location from posts Using a 2010 archive of anonymised tweets, models read a week of posts from 1,697 users and estimated where each lived. Median errors ranged from 20 to 31 km. 135 users were placed within 1 km by at least one model. 70% of those gave themselves away by naming a dorm, a venue or a street. Another 13% were located purely by how they wrote: dialect, slang, sports teams, transit lines. Anthropic says models also tried to de-anonymise users on their own. In one case, a model took a surname from a memorial post about a grandmother and ran a genealogy search. In a separate account-linking test on simulated social media, models assessed around 37,000 words of content, about 2.5 hours of reading for a human analyst, in 11 minutes. ## Drone tests: Opus 5 on top On the weapons side, models wrote flight code for a small simulated drone and revised it after each launch using the flight logs. The drone had no GPS and a single low-resolution camera, and had to hit a vehicle. Against a parked, high-visibility car, Opus 5 struck on 80% of launches, Mythos Preview 70%, Mythos 5 53%, Kimi K3 15% and Sonnet 5 5%. Once the vehicle moved, Opus 5 dropped to 47%. Camouflaged, evading or decoy-surrounded targets were not reliably solved by any model. The notable detail is that Opus 5 beat the newer Mythos models here. Anthropic puts it down to working style. Opus 5 made small edits instead of large rewrites, and it wrote its own small physics model of the drone to test its controller before flying. No other model did that. In a payload-drop test against a moving target, Opus 5 landed 76% of sorties within 5 metres. In the hardest, gusty setting only Opus 5 got meaningful results, at 28%. When GPS was jammed, the frontier models noticed the sensor disagreement and stopped trusting it. Kimi K3 and Sonnet 5 believed the spoofed signal and ended more than 100 metres from the destination. ## Limits and safeguards Anthropic notes that most of the tests are simulations and do not directly measure real-world uplift. But because models worked alone with no internet, it says the results should be read as "a floor rather than a ceiling". Its threat intelligence team has also found real actors misusing models for drone work. Anthropic says it has deployed new classifiers that detect and block weapons-development requests. It adds that open-weights models are not far enough behind the frontier for that gap to count as safety. Source: [Anthropic Frontier Red Team](https://www.anthropic.com/research/intelligence-targeting-conventional-weapons-capabilities?ref=globalfeed.ai).