AI attack surface

Traditional security tools have no visibility into LLM API calls, AI agent tool invocations, MCP server interactions, or model training data access. ManySignal monitors the AI API layer: what data is being sent to AI services, what tools agents are invoking, what system prompts are in use, and whether AI service accounts have permissions consistent with their stated function.

ManySignal monitors AI API request and response payloads for prompt injection signatures — instructions embedded in data fields that attempt to override system prompts or redirect agent behavior. Detection patterns cover direct injection (malicious user input), indirect injection (poisoned data sources), and multi-turn injection (accumulating context manipulation across conversation turns).

Yes. ManySignal monitors access to model training data repositories, model weights storage, and model serving infrastructure. Anomalous access to training datasets (especially by departing employees or external integrations), unauthorized model downloads, and unusual model serving API call patterns are all surfaced as alerts.

ManySignal catalogs all AI service accounts, API keys, and OAuth grants to AI services across the environment. Each AI identity is profiled based on its normal access pattern. When an AI identity accesses resources outside its normal scope, creates new credentials, or exfiltrates data via an AI API call, the anomaly is detected against the established behavioral baseline.

ManySignal monitors MCP server registration events, tool invocation logs, and the data returned to AI agent consumers. Key detection patterns include: an MCP server registered from an unexpected source, a tool invocation accessing resources outside the server's declared scope, data exfiltration volumes through tool call responses, and prompt injection attempts delivered via MCP tool responses back to the orchestrating agent.

No. ManySignal monitors the AI API layer via network tap, cloud provider API audit logs (AWS Bedrock, Azure OpenAI, Google Vertex AI call logs), and API gateway logging — without requiring SDK integration or code changes. For deeper application-layer visibility (request payloads, response content), an optional lightweight SDK is available for Python and TypeScript. The network-level approach covers the majority of security monitoring needs without developer involvement.

Each AI agent identity is assigned a behavioral baseline: expected tool types, normal resource targets, typical call volume ranges, and approved data access patterns. When an agent invokes tools outside its expected set, accesses data beyond its declared context, or exhibits volume patterns inconsistent with its normal operation, ManySignal flags the anomaly. This catches both compromised agents and agents with prompt injection that has redirected their behavior.

EU AI Act requires risk assessment and ongoing monitoring for high-risk AI systems. NIST AI RMF recommends continuous monitoring of AI system behavior. SOC 2 auditors increasingly ask about AI system access controls and monitoring coverage. ManySignal generates evidence packages for AI system monitoring — access logs, anomaly alerts, policy violation records — formatted for regulatory submissions and audit responses.