Prompt injection and tool misuse: the research synthesis
Our synthesis of published prompt-injection and tool-misuse research, with links to the underlying studies.
Assess a team’s workflows, explore policy examples, or start with a reusable template from Alyria (alyria.ai).
Our synthesis of published prompt-injection and tool-misuse research, with links to the underlying studies.
Sourced security metrics with links to primary studies and context for interpreting the results.
Agentic AI maturity is scored on two axes kept separate: how much of a workflow an agent actually runs, and how well its actions are controlled while it runs. The model behind Alyria's five 20-question assessments, with the controls expected at each level.
View the resource →
The measured numbers for prompt injection, wrong tool calls, MCP risk and agent adoption, each linked to its primary source with the date the source carries. Compiled and checkable.
View the resource →
Score an AI coding agent governance RFP on what a vendor can prove on the developer's machine, not on the architecture slide. Every requirement below pairs with a test you can run inside the evaluation window, most of them without the vendor's help.
View the resource →
When agents open 100,000 pull requests, reviewing every diff by hand stops working. A ready-to-adopt secure development lifecycle policy for AI-driven pipelines: normative requirements, roles, the lifecycle phases and their gates, a deterministic risk classifier, and AI-native review. Mapped to SOC 2, NIST SSDF (including the GenAI companion), ISO/IEC 27001 and 42001, and the OWASP Agentic Top 10.
View the resource →
These are reusable resources. Personalized guidance comes from a workflow assessment; observed endpoint inventory comes from the Agent Exposure Report.