AI Security
Safeguard articles tagged "AI Security" — guides, analysis, and best practices for software supply chain and application security.
100 articles
Will AI assistants repeat an insecure example from your docs?
Assistants and search summaries can reproduce code from public docs, including unsafe samples. How to write examples that stay safe when quoted.
Why does LLM-assisted triage still need a deterministic gate?
A model can rank and explain findings, but release decisions need rules that behave the same every time. How to pair the two without confusing them.
Why does an LLM miss a vulnerability disclosed last week?
Models learn from text frozen at a cutoff date. Fresh advisories are missing unless you supply them. How to feed current data into a security prompt.
Why is a CVSS score from an LLM not a risk score?
CVSS measures severity, not risk, and the environment part is yours to fill in. Here is why asking an LLM for one number hides the missing context.
Why must a remote MCP security connector be reachable from the internet?
Claude custom connectors call out from the provider's cloud, so the server must be public or allowlist its IP ranges. What that means for buyers.
Who should own a prompt in production?
Every production prompt needs one accountable owner, a reviewer and a fallback contact. A simple ownership model, a register entry and owner duties.
Who decides what an AI agent may do: the model or the system?
The system must decide, never the model. What complete mediation means for AI agents, where to enforce it, and why prompt rules are not permissions.
Who can add a custom connector on Claude Team and Enterprise?
On Claude Team only Owners add custom connectors, and Enterprise adds custom roles with library access. See who decides, who connects, and what to approve.
Chat assistant, coding agent or scanner: which tool for which job?
A decision guide with twelve common jobs, from explaining an error to finding known flaws in libraries, showing which tool type fits and why.
Which AI coding agent actions should always need human approval?
Require human approval for coding agent actions that are hard to undo, touch secrets or production, change permissions, or reach new network destinations.
Which Claude Code isolation option fits: sandbox, container, VM, or cloud?
Anthropic documents six isolation approaches for Claude Code. A decision guide for choosing between the Bash sandbox, container, VM, and cloud session.
Which AI assistant agent feature should you turn on first?
A rollout order for AI assistant agent features: read-only scheduled tasks, then file generation, desktop apps and remote sandboxes, with exit criteria.
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.