Taiwan Lone Operator Runs 12 Abliterated AI Agents Against Emergency Systems in Seven Taipei Districts

The attacker maintained BOSSAI, a 12-agent cluster with nine active, on a misconfigured server holding 90,000 files of logs, credentials and outputs. Agents scanned hundreds of targets across 11 countries but achieved confirmed access at roughly ten sites, primarily via SQL injection, default credentials and exposed APIs rather than novel exploits. Commercial models refused attack requests, forcing reliance on safety-stripped local instances. Manual handoff of outputs between agents limited scale.

Evidence consists of the operator's own server contents, including dated command histories from July 25 and explicit agent conversations showing failed CVE attempts against a fire-dispatch vendor and successful hardcoded key extraction at an emergency management portal. Taiwan received disproportionate targeting; Mexico yielded 206,800 phone records and Thailand three university account tables. No independent technical attribution or state linkage appears in the dataset.

The pattern shows AI lowering the barrier for persistent low-skill reconnaissance against critical infrastructure while exposing the same infrastructure's reliance on default credentials and public JavaScript secrets. Similar server misconfigurations will likely surface additional BOSSAI-style operations within months.

Taiwanese authorities have not confirmed notification or remediation status. Continued monitoring of exposed emergency APIs and lending firm databases should be prioritized; the 12-day observation window ended without evidence of lateral movement beyond initial data pulls.