Skip to main content

AgentGuard — AI Runtime Security

AgentGuard is BleedWatch's runtime security module for AI agent workloads. It monitors AI agents executing in your infrastructure in real-time, detects anomalous behavior, and can automatically terminate compromised sessions.

Separate Product Module

AgentGuard is an add-on module available on Enterprise plans. Contact your account representative to enable it.

Three Security Layers​

1. eBPF Monitoring​

AgentGuard deploys an eBPF-based daemon that instruments AI agent processes at the kernel level. It observes:

What It MonitorsWhy It Matters
Process executionDetects unexpected child processes spawned by agents (e.g., shells, network tools)
Network connectionsIdentifies data exfiltration attempts — outbound connections to unknown hosts
File system accessAlerts on agents reading sensitive files outside their sandbox (credentials, configs, SSH keys)
System callsDetects privilege escalation attempts or unusual syscall patterns

eBPF monitoring operates at the kernel level with minimal performance overhead (under 2% CPU) — no agent code modification required.

2. Anomaly Detection​

An AI behavioral analysis engine profiles each agent's normal activity and flags deviations:

  • Prompt injection detection — Identifies when an agent's behavior diverges from its intended instructions, suggesting a prompt injection attack
  • Tool abuse — Detects agents making excessive or unusual API calls (e.g., mass data queries, credential access)
  • Data exfiltration patterns — Recognizes when an agent attempts to encode, compress, or transmit large volumes of data to external endpoints
  • Lateral movement — Flags agents that attempt to access resources or services beyond their authorized scope

3. Kill Switch​

When a threat threshold is exceeded, AgentGuard can automatically:

  • Terminate the agent session immediately
  • Revoke the agent's API tokens to prevent further access
  • Isolate the container from the network
  • Generate a forensic snapshot of the agent's state at termination time

The kill switch can operate in two modes:

ModeBehavior
AutomaticTerminate immediately when threat score exceeds threshold
ManualAlert the operator and await confirmation before termination

Daemon Status​

The AgentGuard page displays the current daemon status:

  • Running — The eBPF daemon is active and monitoring
  • Unavailable — The daemon is not deployed or not responding
  • Version — Current daemon version and uptime

Deployment​

AgentGuard requires deploying the agentguard-daemon container alongside your AI agent workloads. The daemon:

  1. Runs as a privileged container with access to the host kernel (required for eBPF)
  2. Communicates findings back to BleedWatch via a secure gRPC channel
  3. Requires Linux kernel 5.8+ with BTF (BPF Type Format) support

Detailed deployment instructions are provided during Enterprise onboarding.

Use Cases​

Securing Code-Generation Agents​

AI agents that generate and execute code (like coding assistants) can be manipulated to run malicious commands. AgentGuard monitors for:

  • Spawning unexpected processes (curl, wget, nc, reverse shells)
  • Accessing credential files (.env, ~/.aws/credentials, SSH keys)
  • Making outbound connections to unauthorized endpoints

Protecting Data-Analysis Agents​

AI agents with database access can be tricked into exfiltrating data. AgentGuard detects:

  • Unusually large query result sets
  • Data encoding patterns (base64, compression) in output
  • Outbound data transfers that exceed baseline

Multi-Agent Orchestration​

In multi-agent systems, a compromised agent can attempt to manipulate other agents. AgentGuard monitors inter-agent communication for:

  • Prompt injection payloads in agent-to-agent messages
  • Unauthorized tool invocations
  • Scope escalation beyond assigned roles