AgentGuard Research
Practical guidance, independent reviews, and clear comparisons for teams building and securing AI agents.
An agent can make a bad security decision before it ever calls a tool: it can accept a package, plugin, skill, tool description, or MCP server that is not what its owner believes it to be.
By Agent Guard Team4 min read
Read articlePrompt injection lets untrusted text override an AI system’s intended instructions. It may come directly from users or indirectly through webpages, documents, emails, and tools. Defenses should separate data from authority, restrict permissions, validate actions externally, require approval for sensitive operations, and log decisions. Simple phrase filtering is insufficient protection.
By Agent Guard Team6 min read
Read article