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.
The Claude API code execution tool runs Python and bash in a container with no internet, fixed resource limits, and 30 day retention. What that means for you.
Sort every AI agent tool by what its verb can change: read, write, delete or send. A four row risk table, with the default control for each class.
A downloaded model can ship a Jinja chat template that shapes every prompt. What it is, where it lives and what to check before you use one.
AI answers about SAST are useful for definitions and question lists, but weak on current features and fit. Which claims to verify and where to check them.
No. PyTorch says distributed checkpoints are for your own trusted jobs, and weights_only protections do not apply to them. How to handle shared ones.
Yes. Text filled into a template can change instructions, and template engines add risks too. How to separate data from instructions and what to test.
Not reliably. A backdoor behaves normally until a trigger appears, so sample tests can miss it. Controls on data, provenance and permissions matter more.
Yes, with limits: a narrow write tool, human confirmation, input validation, rate limits and logging, as the MCP spec advises. A practical setup guide.
An assistant can answer from internal security docs if they are current, access aware and cited. These are the checks to make before turning one on.
Weekly insights on software supply chain security, delivered to your inbox.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.
