Sysdig's threat research team found that in November 2025 attackers used LLMs to complete a full AWS attack chain in under eight minutes — from credential theft to full administrative control — compressing the cloud attack lifecycle from hours to minutes.
The attack started from valid AWS credentials exposed in a public S3 bucket. The associated IAM user held Lambda and limited Bedrock access; the attacker enumerated services and injected LLM-generated malicious code into an over-privileged Lambda function to escalate to admin keys. The root cause remained a basic hygiene failure: exposed credentials. With admin access, the attacker moved laterally across 19 AWS principals, hijacked LLM calls to Bedrock models, and attempted high-end GPU instance abuse — some code failing due to LLM hallucinations.
Experts note no new techniques were involved: AI removed attacker hesitation, compressing reconnaissance and escalation into rapid sequences that eliminate defender buffer time. Sysdig recommends least-privilege principles, restricting Lambda permissions, keeping sensitive S3 buckets private, enabling Lambda versioning and Bedrock invocation logs, and monitoring mass-enumeration behavior.