Claude is getting unnervingly good at finding targets

For years, the intelligence community relied on a simple form of security through obscurity: finding a needle in a digital haystack was just too expensive for anyone without a nation-state budget.

Anthropic’s Frontier Red Team just blew that assumption wide open with a chilling new evaluation of how cutting-edge AI handles tactical intelligence targeting and conventional weapons tasks.

Beating the world's best at real-world hide and seek

Anthropic tested its unreleased frontier models alongside open-weights rivals across thousands of simulated national security workflows.

The standout benchmark evaluated photo geolocation. Anthropic fed models 6,000 outdoor images stripped of all metadata, forbidding reverse image search or any external tools.

To see how models stacked up, researchers compared them against top competitive GeoGuessr players. Human champions in the top 0.01% typically land within a median error of 151 km. Claude Mythos Preview and Mythos 5 shattered that benchmark, clocking median errors of 37.0 km and 47.2 km, respectively. Roughly 23% of their guesses were within just 1 km of the actual spot.

Chinese open-weights models are not far behind. Moonshot AI's Kimi K3 scored a median error of 385 km, but placed 16.7% of its predictions within 1 km, roughly 1.7 times the pinpoint accuracy of Claude Sonnet 5.

The models were equally ruthless at cross-platform de-anonymization. When tasked with analyzing 37,000 words of simulated chat logs across Telegram, WhatsApp, Facebook, and Instagram to link burner accounts, Mythos Preview took just 11 minutes to complete an investigation that would take a trained human analyst hours.

Lowering the cost of the kill chain

In intelligence work, the hardest step of the kill chain is often the very beginning: finding and fixing targets. Huge troves of personal data have always been cheap or freely available online, but the analyst hours required to correlate it served as a natural barrier.

Anthropic notes that AI has effectively collapsed that cost barrier. Small adversarial groups can now conduct surveillance operations that previously demanded a fully staffed spy agency.

Anthropic says it has already deployed new classifiers to block users from weaponizing Claude for targeting and drone design. However, the team explicitly warned that open-weights models are rapidly closing the gap, meaning platform-level guardrails alone cannot stop the shift.

Why it matters

This is a stark reminder that frontier AI risks are no longer confined to sci-fi scenarios like rogue bioweapons; they are already transforming conventional tactical warfare and surveillance.

For everyday internet users, the take-home message is uncomfortable. Scrubbing EXIF data from your vacation photos no longer protects your location. If a multi-modal model can deduce your street corner based on a curb design and tree bark, personal operational security has essentially evaporated.

From an industry perspective, we think Anthropic's public transparency here is both strategic and necessary. By publishing hard numbers rather than vague safety warnings, it forces regulators and rival labs to look beyond cybersecurity benchmarks.

The glaring dilemma, though, is the open-weights cat is already out of the bag. You can patch Claude all day, but when models like Kimi K3 show comparable targeting capabilities without Anthropic-style API filters, platform moderation becomes a speed bump rather than a shield.