AI Security
What an AI agent can do with your logged-in session
When an agent works inside your browser or desktop, it inherits everything you can do. A plain explanation of ambient authority and how to shrink it.
Oct 18, 20263 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
A vendor turned on an AI feature by default: what do you check?
When a vendor turns on an AI feature by default, check data flow, scope, training use, admin controls and contract terms in a week, then decide.
Oct 18, 20263 min read
AI Security
Two MCP servers expose the same tool name: what happens?
Tool names are unique only within one MCP server. When two servers both offer search, clients need a prefix or policy. Here is the risk and the fix.
Oct 18, 20263 min read
AI Security
Twelve test cases to run before an AI agent goes live
A concrete starter set of agent tests covering scope, injection, refusals, loops, secrets and shutdown. Run each several times and record the results.
Oct 18, 20263 min read
AI Security
Training data poisoning: controls a data team can run this quarter
Poisoned training data changes model behavior. Practical controls for data teams: source trust, snapshots, review of changes and anomaly checks.
Oct 18, 20263 min read
AI Security
Train test data leakage: a correctness bug and a governance signal
Fitting preprocessing on all your data inflates model scores. Why leakage matters to data governance, with the documented example and a fix.
Oct 18, 20263 min read
AI Security
Claude thumbs up and down: the whole chat is kept for 5 years
Pressing a feedback button in Claude stores the related conversation for up to five years and may feed training. How to give feedback without oversharing.
Oct 18, 20263 min read
AI Security
Six design patterns for agents that resist prompt injection
A 2025 research paper names six agent design patterns that limit prompt injection by structure. Here is what each does and what it costs in usefulness.
Oct 18, 20264 min read
AI Security
Should you scan training and eval data for secrets and personal data?
Yes. Scan datasets before training, fine-tuning or evals for keys, personal data and poisoned records, because models can repeat what they were given.
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.