AI AGENT SECURITY

Secure your AI agents

Validate every tool call before execution. Detect privilege escalation, monitor agent behavior, and block dangerous commands across OpenClaw, LangGraph, CrewAI, and more-all in real-time.

Start for Free View documentation

THE CALL, AND THE RETURN LEGfig. 01

The exposure is the return leg. Tool output is content nobody on your side wrote, and it is scanned before it re-enters the agent's context — structurally first, with no model call, and by an LLM judge only when the structural scan is clean.

What is checked, and when

Tool Call Validation: Intercept and validate every tool call before execution. Block dangerous operations like file deletion, shell commands, and unauthorized API calls.
Privilege Escalation Detection: Detect and block attempts to escalate privileges, such as agents trying to modify permissions or access restricted resources.
Behavior Monitoring: Track agent behavior over time to establish baselines and detect anomalies like unusual tool usage patterns or rapid-fire operations.
Human-in-the-Loop: Require human approval for sensitive operations like sending emails, modifying user data, or executing financial transactions.
Session Isolation: Isolate agent sessions to prevent cross-contamination and limit the blast radius of compromised agents.
Real-time Alerts: Get instant notifications when suspicious agent behavior is detected. Integrate with Slack, PagerDuty, and webhooks.
Tool Injection Detection: Detect and block attempts to inject malicious tool calls or manipulate agent tool usage through crafted prompts targeting function-calling workflows.
Jailbreak Detection (LLM): LLM-powered jailbreak detection catches sophisticated bypass attempts that evade traditional pattern matching, including multi-turn, encoded, and role-play attacks.
OpenClaw Agent Defense: Purpose-built protection for OpenClaw agents: defend against indirect prompt injection and tool-hijacking exploits, and validate skill marketplace packages.
Multi-Turn Intent Drift Detection: Detect crescendo attacks and slow-burn manipulation across multi-turn conversations. Identifies gradual intent drift before agents are compromised.
Agent Identity Registry: Register agents and issue a per-agent credential (pgag_, bcrypt-hashed at rest, shown once), rotatable and revocable through the API. Requests today carry a self-asserted agent ID; credential checking on the request path is on the roadmap, not yet enabled.
Behavioral Drift Detection: Establish behavioral baselines using tool-usage distributions and detect drift via Jensen-Shannon divergence. Alerts fire when an agent's behavior shifts beyond the configured threshold.
Tamper-Evident Audit Trail: Every audit event is SHA-256 hash-chained to its predecessor, forming a cryptographic append-only chain. Verify chain integrity over any time range with a single API call.
Content Safety Classification: Classify harmful intent using an LLM safety classifier. Detect violent, sexual, self-harm, and hateful content before it reaches your agents or users.
EU AI Act & ISO 42001 Aligned: Technical controls mapped to EU AI Act Articles 9-15 and ISO/IEC 42001 Annex A. Risk management, automated record-keeping, transparency, human oversight, and governance reporting, built in, not bolted on.
Framework Integrations: Native support for LangChain, LangGraph, CrewAI, OpenClaw, Pydantic AI, OpenAI Agents SDK, and Vercel AI SDK. Drop-in security for any agent framework.

How AI agent security works

01
Intercept
Every tool call from your AI agent passes through PromptGuard's security layer before execution.
02
Validate
We validate the tool, arguments, and context against your security policies and behavioral baselines.
03
Execute or Block
Safe operations execute immediately. Dangerous or suspicious calls are blocked or require human approval.

Simple integration

python

from promptguard import PromptGuard

pg = PromptGuard(api_key="your-api-key")

# Validate a tool call before execution
result = pg.agent.validate_tool(
    agent_id="agent-123",
    tool_name="execute_shell",
    arguments={"command": "rm -rf /"}
)

if result.action == "block":
    print(f"Blocked: {result.reason}")
    # Risk level: {result.risk_level}
else:
    # Safe to execute
    execute_tool(result.tool_name, result.arguments)

Why PromptGuard for AI agents?

Included in all plans (even free tier)
Real-time tool call validation
Behavioral anomaly detection
Human-in-the-loop workflows
Self-hostable with open-source SDKs

Other controls on the same path

Red Team Testing

Test your agent security with automated adversarial attacks.