An AI-Assisted AWS Break-In in 8 Minutes (Feb 2026): From Public S3 Credentials to Admin
In early February 2026, researchers described an AWS intrusion in which an attacker went from stolen credentials to administrative privileges in under ten minutes, with heavy use of AI-assisted automation, according to reporting by The Register.
What was reported
- Initial access: the attacker found valid test credentials in publicly accessible S3 buckets. They belonged to an IAM user with read and write permissions on AWS Lambda and limited permissions on Amazon Bedrock.
- Escalation: using those permissions, the attacker escalated to administrative access within minutes. Lambda write access is a well-known escalation path when functions run with more privileged execution roles.
- Spread: the attacker ultimately compromised 19 distinct AWS principals.
- Abuse: they used Bedrock models and GPU compute, the kind of resource abuse that turns into large bills quickly.
- Speed: researchers attributed the pace to AI-assisted tooling that automated reconnaissance and decision-making.
Why it matters
The individual weaknesses weren't new: credentials in public buckets, over-permissioned test users, Lambda-based privilege escalation. What changed was speed. Many organizations' detection and response processes are measured in hours. An attacker who reaches admin in eight minutes and spreads across 19 identities shortly after leaves almost no window for human response.
What to do now
- Find and remove credentials from S3 buckets and other storage, and block public bucket access account-wide.
- Eliminate long-lived IAM user keys, especially for test accounts.
- Review Lambda and Bedrock permissions: who can create or update functions, and which roles can those functions assume?
- Automate first response: for high-confidence findings, revoke credentials automatically instead of waiting for an analyst.
- Set budget and Bedrock usage alarms.
The bigger picture
AI is now part of attackers' toolkits. Defenders don't need AI to respond — they need fewer exploitable paths and faster automated containment.