AI Security
Safeguard articles tagged "AI Security" — guides, analysis, and best practices for software supply chain and application security.
100 articles
What changed in the 2026-07-28 MCP specification for buyers?
MCP 2026-07-28 is stateless, adds server discovery and deprecates sampling and protocol logging. What that means when you evaluate connectors.
What can go wrong with an AI agent? Five kinds of failure explained
Five ways an AI agent can fail, from honest mistakes to hidden instructions, with a plain example of each and the first control to try for it.
What can an AI desktop app see on your computer? A permission checklist
An AI desktop app may read files, the clipboard, the screen and run commands. A checklist to map each reach, limit it and verify it before rollout.
What can an AI coding assistant see and change on my computer?
A plain guide to what AI coding tools can read, edit and run on your machine, how permission modes and sandboxes limit them, and what to check in settings.
What does an Ollama Modelfile control, and what should you review?
A shared Modelfile sets the weights source, system prompt, example messages, template and parameters. What each instruction does and what to check.
Which vulnerability decisions should an LLM never make alone?
Models can draft and explain, but some calls need a named person: risk acceptance, exceptions, severity overrides and release blocks. A decision map.
Is multi-node vLLM traffic encrypted? What network isolation needs
vLLM's docs say multi-node traffic is insecure by default, and PyTorch Distributed has no auth or encryption. Here is how to isolate the nodes.
Which vLLM flags and variables should never be on in production?
Dev mode, profiler endpoints, runtime LoRA updates and the tokenizer info endpoint each add risky routes. What vLLM's docs say to leave off or guard.
Does the vLLM API key protect every endpoint? No, here is the gap
No. vLLM's API key covers some path prefixes, but its own security docs list inference and control endpoints outside them. Here is how to close the gap.
How do you stop SSRF through multimodal inputs on a vLLM server?
Set an allowlist of media domains, block redirects and cap download sizes. The vLLM docs explain why URL inputs to an inference server become an SSRF risk.
How do you validate an AI-generated patch without trusting the AI?
Treat AI output as untrusted input. Validate patches with deterministic checks on paths, packages, size and tests before a human reads the diff.
Should you trust an AI-suggested security fix? Treat it as untrusted
Treat an AI-suggested fix like a patch from an unknown contributor: read it, test it, rescan it. A step list for verifying it before merge.
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.