Agent-Specific Threats
AI agents face threats that traditional applications don’t:1. Prompt Injection → Credential Theft
Attack: Adversarial input tricks the LLM into callinghexr_tool() with attacker-controlled parameters, exfiltrating data.
Hexr Mitigation:
- LLM Guard scans all prompts for injection attempts
- OPA policies restrict which services each role can access
- Credential scoping — even if tricked, the agent can only access its authorized services
- Short-lived credentials — stolen tokens expire in 15 minutes
2. Tool Confusion → Unauthorized Access
Attack: Agent is manipulated into calling the wrong tool or accessing unauthorized resources. Hexr Mitigation:- SPIFFE identity — OPA verifies the specific process identity, not just the pod
- Gateway validation — tool calls are validated against the agent’s registered capabilities
- Audit trail — every tool call is logged with full SPIFFE context
3. Agent-to-Agent Manipulation
Attack: A compromised agent sends malicious tasks to other agents. Hexr Mitigation:- mTLS — all A2A communication is SPIFFE-authenticated
- OPA policies — controls which agents can communicate
- Task validation — A2A sidecar validates message schema
4. Secret Exfiltration
Attack: Agent code or LLM output leaks stored secrets. Hexr Mitigation:- Vault scoping — secrets are per-agent, per-role
- PII scanner — LLM Guard catches secrets in LLM output
- No env vars — secrets never in environment variables
5. Code Execution Escape
Attack: LLM-generated code executed by the agent escapes the sandbox. Hexr Mitigation:- Firecracker microVM — hardware-level isolation
- No network — sandbox has no outbound connectivity
- Resource limits — prevents resource exhaustion
- Destroyed after use — no persistent state