AI Security
Safeguard articles tagged "AI Security" — guides, analysis, and best practices for software supply chain and application security.
100 articles
Where does the risk live in a model repo: weights, code or config?
A model repo holds weights, code, tokenizers and config. Each carries a different risk. Here is a map of what to check in each file type before you load.
Where did your few-shot prompt examples come from? A provenance check
Few-shot examples in a prompt teach the model by demonstration. Check where they came from, what they contain and who can change them before you ship.
What should you log from a cloud coding agent session?
A cloud agent session should leave a trail: who started it, which commands ran, which hosts were blocked, and what changed. Fields worth keeping.
What should you log about prompts and datasets for an AI incident?
Log versions and identifiers, not raw secrets: template version, model version, dataset hashes and retrieved document IDs. A short list for fast triage.
What are split view and frontrunning dataset poisoning attacks?
Two practical ways to poison web scale training data, per a 2023 paper: change content after indexing, or edit just before a snapshot. Defenses explained.
What should a platform do when an AI agent adds a dependency?
AI coding agents add dependencies fast. Treat each one like a human change: check it in the pull request, apply policy, and record who approved it.
What scopes should a security MCP connector ask for?
A security connector should start with read scopes and ask for more only when needed. What least privilege looks like and what to do when it asks too much.
What does PyTorch's security policy count as a vulnerability?
PyTorch says secure use is your job: loading untrusted models is like running untrusted code. Here is what the policy calls a bug and what you should do.
What access does a security vendor's AI agent hold in your environment?
Before connecting a vendor's AI agent, list its identity, scopes, tokens, and actions. Ask for least privilege, short lived access, and logs you can read.
What is Keras safe_mode, and when is it OK to turn it off?
Keras safe_mode blocks unsafe lambda deserialization when loading .keras models. Here is what it does, what turning it off risks and a safe way to proceed.
What is an MCP guardrail and what does it do, in plain language?
An MCP guardrail is a checkpoint between an AI agent and its tool servers that applies your rules to each call. Here is how it works and what it cannot do.
What is an MCP connector? Host, client and server for buyers
An MCP connector is a server that gives an AI app tools and data. This explainer defines host, client, server and tool in plain words for security buyers.
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.