AI Security
Turning AI red team findings into regression tests
How to convert each successful AI attack into a repeatable regression test, so fixes stay fixed when prompts, models or data sources all keep changing.
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
trust_remote_code=True: what actually runs on your machine
trust_remote_code=True lets a Hub repository supply Python modeling code that runs locally. Here is when you need it and how to review it first.
Oct 18, 20263 min read
AI Security
torch.load weights_only: what it does and does not cover
weights_only=True limits what torch.load will unpickle to tensors and simple types. Here is what that buys you, and where you still need to be careful.
Oct 18, 20264 min read
AI Security
Tool permissions vs guardrails: which control does what
Permissions limit what an agent can do. Guardrails check what it tries to do. A decision guide to using both, grounded in OWASP and MCP guidance.
Oct 18, 20263 min read
AI Security
Tool gating or Guard: which MCP control do you need?
Tool gating limits which Safeguard tools an assistant sees. Guard enforces live policy on MCP traffic in general. They are independent. How to choose.
Oct 18, 20263 min read
AI Security
Token exchange for AI agents: delegation vs impersonation
OAuth token exchange (RFC 8693) lets an agent act for a user without becoming them. Learn the act claim, may_act, and why delegation beats impersonation.
Oct 18, 20263 min read
AI Security
Testing AI systems under the NIST MEASURE function
What the NIST MEASURE function expects: documented test sets, production monitoring, safety, security, privacy and bias checks, and independent review.
Oct 18, 20263 min read
AI Security
Stop sign stickers and physical adversarial examples explained
Researchers fooled a road sign classifier with stickers on a real stop sign. What the study showed, what it did not, and how builders test for it.
Oct 18, 20264 min read
AI Security
Stop AI agents leaking sensitive data through their context
OWASP's sensitive information disclosure guidance applied to agents: sanitize, limit data access, redact, and never trust a prompt to hold secrets.
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.