“Confused Deputy” in third-party SaaS integrations: when the problem isn’t your account, but the intermediary

by

The “Confused Deputy” pattern appears when a SaaS with access to your cloud does not correctly validate which customer is requesting an action. The result can be that another account within the SaaS itself—or an attacker—ends up operating in your tenant. What it looks like in real integrations (CI/CD, monitoring, billing) and how to mitigate it with concrete guardrails (ExternalId, OIDC with iss/aud, conditions by tags/IP) without breaking operations.

Indirect Prompt Injection and credential theft in RAG architectures: how a “harmless” document ends up exfiltrating your cloud

by

In RAG architectures, a PDF can become a “control channel” for your agent: it injects instructions the model obeys and, if the runtime has broad permissions (S3/DB/email), the real consequence is data and credential exfiltration. This article breaks down the abuse, the architectural failure, and how to harden permissions and action flows without breaking the product.

The Key Vault trap: the set-and-forget syndrome

by

Using a secrets manager does not “buy” you security automatically. The set-and-forget syndrome appears when it is assumed that Azure Key Vault, AWS Secrets Manager, or GCP Secret Manager solve encryption, access control, and rotation by default. In practice, failures come from massive permissions, non-existent rotation, and a false sense of network protection.

The Real Danger of “Public Buckets” in Data Analytics and Machine Learning

by

In analytics and ML it is common to dump datasets into “temporary” buckets to speed up experiments. When those buckets become public (read or, worse, write), the risk goes from PII leakage to data manipulation and infrastructure abuse. This operational guide explains how these blind spots are created, how to detect them, and how to close the exposure without breaking pipelines.

From EKS/AKS to the root account: the risk of privilege escalation from a compromised Pod

by

A compromised Pod should not be able to “jump” from the cluster to the cloud. In practice, a bad combination of identities (IRSA/Workload Identity), excessive permissions, and lack of isolation can turn an app vulnerability into full tenant control. This article walks through how it happens, what signals to look for, and how to harden it without slowing operations.

Break-glass accounts: necessary… and dangerous if nobody controls them

by

Break-glass accounts exist to regain control when everything fails. The problem: if they are created “just in case” and nobody governs them, they end up being the shortest path to internal abuse, lateral movement, and control evasion. This guide makes it practical what they are, how they are abused in real life, and how to operate them without turning them into a permanent backdoor.