AI Security
Keras safe_mode and Lambda layers: why model files can run code
Keras load_model has a safe_mode flag that blocks unsafe Lambda deserialization in v3 files. Here is what it covers and why you should leave it on.
Oct 18, 20264 min read
Resources
Supply Chain Security, in plain English.
Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
Filter
Articles
RSS feedAI Security
Jailbreak or prompt injection: which term fits your incident?
Jailbreaking is a form of prompt injection aimed at removing safety rules. A short guide to choosing the right term in tickets, reports and audits.
Oct 18, 20263 min read
AI Security
Is the NIST AI RMF mandatory? What voluntary really means
The NIST AI RMF is voluntary and not a certification, but contracts, customers or other rules may still point to it. How closely should you follow it?
Oct 18, 20263 min read
AI Security
Is an LLM judge guardrail enough to stop prompt injection?
A second model that screens inputs and outputs is useful, but it can be injected too. Why guardrail models are one layer and what to pair them with.
Oct 18, 20263 min read
AI Security
How to inspect a pickle model file without running it
Use pickletools to list the imports and opcodes in a pickle model file without executing it, and learn which opcodes matter when you review one.
Oct 18, 20263 min read
AI Security
Input or output moderation: where to filter LLM content
Should a content filter check what users send, what the model returns, or both? A decision guide for AI products with real users, with the trade-offs.
Oct 18, 20263 min read
AI Security
Hypothesis-driven threat hunting with an AI assistant
How to use an LLM in a threat hunt without losing rigor: write the hypothesis first, let the model draft queries, and verify every result against raw data.
Oct 18, 20263 min read
AI Security
Human approval for LLM tool calls: what the approver must see
Human in the loop only works if the approver sees the exact action. What to show, which actions to gate, and how to avoid rubber stamp fatigue.
Oct 18, 20263 min read
AI Security
Hugging Face scan results: what a clean badge does and does not mean
The Hub scans uploads with ClamAV, a pickle import check and partner scanners. A clean result lowers risk but is not proof a model is safe to load.
Oct 18, 20263 min read
Stay informed
Weekly insights on software supply chain security, delivered to your inbox.
Self-healing security runs on Safeguard.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.