AI Security
Safeguard articles tagged "AI Security" — guides, analysis, and best practices for software supply chain and application security.
100 articles
Training data, retrieval corpus or prompt: which carries which risk?
An LLM app takes in text through training data, a retrieval corpus and prompts. Each fails differently and needs a different control. A comparison table.
Is a model you pulled for a prototype part of your supply chain?
Yes. Pre-trained models and datasets pulled for a prototype are supply chain inputs. Here are the checks to run before they reach production.
Tabletop script: your AI assistant installs a fake package
A 45 minute tabletop script for a coding assistant that recommends a nonexistent package an attacker then registers: detection, scope, policy fixes.
Tabletop script: a connected MCP server starts misbehaving
A 45 minute tabletop script for an MCP server that changes behaviour after an update: find who connected it, limit tool access, review logs, set approvals.
Tabletop scenario: an AI coding agent installs a lookalike package
A 60 minute tabletop where an AI coding agent installs a near match package: injects, questions on tokens and sandboxes, and what good answers look like.
Should you log an AI agent's reasoning text in the audit trail?
Reasoning text is not a reliable record of intent. Log actions, approvals and outcomes in the audit trail, and keep reasoning in short-lived debug logs.
Load prompts from a database or keep them in the build?
Keep prompts in the build by default. Runtime loading lets anyone with database access change behavior without review. When each option makes sense.
Should you assign vulnerability alerts to an AI agent?
Platforms now let you assign alerts to AI agents that open draft fix PRs. Here is what that changes in the exposure window and which controls to keep.
Should we ban shadow AI or run an amnesty first?
When a ban works, when an amnesty finds more, and how to run a two week disclosure window that ends in an approved list. A guide for engineering leaders.
Should research mode be allowed to call security connector tools?
Research runs can call connector tools without more approvals. Turn off write tools first and keep sensitive reads narrow. A guide for security teams.
Is it safe to paste an SBOM into an LLM chat?
An SBOM looks harmless, but it can reveal internal names, versions and structure. Here is what to check before sharing one with an AI assistant.
Should an AI agent be allowed to suppress security findings?
An AI agent can draft a suppression with evidence, but a human should approve it. What agents may propose, what they may not do, and how to audit them.
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.