AI Security: Lookups
The AI Security module uses 11 lookups in two categories: attack signatures and baseline values.
Intended audience: information security analysts and SOC engineers
Signature Lookups
These lookups are used by jobs to detect threats in requests, responses, and agent actions.
| Lookup | Purpose |
|---|---|
prompt_injection_patterns | Signatures of prompt injection attempts: instruction overrides, role substitution, system prompt extraction, and jailbreaks |
llm_warning_keywords | Data leakage signatures: access tokens and passwords, personal data, NDA markers, and indicators of web attacks (XSS, SQL injection, and shell injection) |
system_prompt_patterns | Markers associated with model system prompts |
model_storage_paths | Paths and file extensions for model artifacts, including weights and tokenizers |
training_data_paths | Paths and file formats for training datasets |
dangerous_agent_actions | Signatures of dangerous commands and actions, including destructive operations and the download and execution of third-party code |
agent_process_names | AI agent runtime process names used to associate a detected action with a specific agent |
Baseline Lookups
These lookups are used by jobs to detect deviations from expected behavior and serve as allowlists or baselines.
| Lookup | Purpose |
|---|---|
agent_permission_baseline | Baseline (expected) AI agent permissions: approval mode, network access, allowed prefixes, sandbox mode, and origins |
model_storage_authorized | Users authorized to access model repositories |
training_data_authorized | Users authorized to access training data |
agent_known_destinations | Network destinations allowed for AI agents |
Important
Lookups containing deployment-specific data must be adjusted for the target deployment.