AWS Security Digest #260 -
This digest curates several technically useful AWS security items rather than presenting a single original finding. The strongest material covers experiments showing LLM agents can turn a leaked AWS access key into S3 data exfiltration in about a minute by enumerating IAM, retrieving cross-account credentials from a staging bucket, assuming a higher-privilege role, and then locating and downloading target objects. It also highlights two defensive approaches for AI agents using AWS: an iam-agent-proxy design that gives agents fake access keys, validates and re-signs SigV4 requests with real credentials while logging resolved IAM actions, and a Unix-socket credential broker model that keeps short-lived STS credentials in memory instead of env vars or ~/.aws files. The issue additionally notes AWS response guidance for the branded Linux privilege-escalation bug “Copy Fail,” including AL2023 livepatch availability and the continued exposure of self-managed EC2, ECS-on-EC2, and long-tail unmanaged Linux fleets.