This week's articles
Hacking Gemini: A Multi-Layered Approach
The article describes exploiting multi-layered architecture discrepancies in Gemini to bypass Markdown sanitization. The researcher achieved image injection through linkification quirks and context bridges (Gemini-to-Colab), enabling workspace data exfiltration via indirect prompt injection despite existing protections.
#ai
#attack
#saas
#gcp
GitHub Actions Security: A Guide to Common Risks
This article examines GitHub Actions security risks including secrets management failures, overprivileged tokens, dependency pinning issues, supply chain attacks via compromised actions, script injection, workflow injection through pull requests, artifact poisoning, and self-hosted runner compromise.
#ci/cd
#attack
#supply-chain
The Sins of Security Vendor Research
A post which critiques security vendor research practices, identifying four common pitfalls: using fear-based marketing tactics, falsely claiming novelty while ignoring prior work, making unsupported statistical correlations, and compromising editorial integrity for marketing purposes.
#process
#strategy
Towards a secure by default GitHub Actions
GitHub announces security changes to pull_request_target event and environment protection rules, effective December 8, 2025, to reduce vulnerabilities and make Actions more secure by default.
#announcement
#ci/cd
#defend
Public Report: Google Private AI Compute Review
NCC Group conducted a 100 person-day security review of Google's Private AI Compute system across two phases, evaluating architecture, cryptography, attestation, IP-blinding relay, T-Log system, and frontend components to ensure cloud-based AI processing maintains local-only privacy guarantees.
#ai
#gcp
#attack
#defend
|