Prompt-injection testing
Direct injection (OWASP LLM01), indirect injection via RAG corpus, tool-result injection. Probe families align to OWASP LLM Top 10 (2025).
LLM Security Posture Management is a focused layer within AI Security Posture Management (AI-SPM) that helps teams test, monitor, and govern LLM applications, prompt flows, model endpoints, and runtime controls.
AI-SPM covers the full AI control surface (assets, agents, MCP servers, RAG, vector DBs, gateways). LLM Security focuses the lens on the language-model-specific surface: prompts, completions, model endpoints, and inline filtering.
Direct injection (OWASP LLM01), indirect injection via RAG corpus, tool-result injection. Probe families align to OWASP LLM Top 10 (2025).
Insecure output handling (LLM02). Detects responses that leak credentials, internal URLs, or executable content; flags downstream consumers without sanitisation.
Probes for inadvertent leakage of training data, system prompts, embedded credentials, or context-window contents (LLM06). 48 DLP patterns + custom regex.
Tests for over-broad tool authorisations and missing human-in-the-loop on destructive operations (LLM08 + ASI Agentic Top 10).
Optional self-hosted DLP firewall + tool allowlist + per-domain budgets. Prompts never leave the customer VPC when the gateway is in use.
Each adversarial response scored by three independent judges (Anthropic, OpenAI, Google) plus a meta-judge. Reduces single-model bias and surfaces low-confidence cases.
LLM01 Prompt injection
LLM02 Insecure output handling
LLM03 Training data poisoning
LLM04 Model DoS
LLM05 Supply chain
LLM06 Sensitive info disclosure
LLM07 Insecure plugin design
LLM08 Excessive agency
LLM09 Overreliance
LLM10 Model theft
Auditors verify our control mapping against the same documents we read. Each item below points to the canonical publication.
Last reviewed:
Most teams start with LLM Security testing and expand to full AI-SPM as agents, MCP servers, RAG pipelines, and runtime gateways come into scope.