A Single Cloud Compromise Can Feed an Army of AI Sex Bots

Permiso Security researchers leaked a test AWS key and observed attackers quickly using the stolen Bedrock access to power subscription AI sex-chat services that employ jailbreak prompts to bypass model restrictions; their two-day honeypot recorded ~75,000 model invocations and generated a $3,500 bill, with several prompts indicating illegal sexual content including child exploitation. The report highlights widespread risks from exposed cloud credentials, gaps in default logging of LLM prompts/outputs, and ongoing attacker techniques to detect and avoid accounts with logging enabled, while vendors like AWS and Anthropic describe mitigations and policy work.