Security & AI Agent Threats

Cheqpoint is a security control that sits between your AI agents and the real world. This page covers the AI agent threat landscape, how Cheqpoint addresses each threat, and our own infrastructure security posture.

Kill switch - suspend an agent instantly

bash

# Dashboard: AI Assistants → <agent> → Kill Switch (toggle off)
# All pending and incoming requests from this agent are immediately suspended.

# Via API (e.g. from a Falcon Fusion playbook or SIEM alert):
curl -X POST https://app.cheqpoint.co/api/agents/<agentId>/kill \
  -H "x-api-key: <connection-key>" \
  -d '{ "reason": "Suspended - prompt injection suspected" }'

Verifying webhook signatures

javascript

import crypto from 'crypto';

// Verify that events sent to your Notification URLs came from Cheqpoint
const signature = crypto
  .createHmac('sha256', process.env.CHEQPOINT_WEBHOOK_SECRET)
  .update(JSON.stringify(payload))
  .digest('hex');

if (signature !== headers['x-cheqpoint-signature']) {
  throw new Error('Unauthorized');
}

CEF export for SIEM forensics

text

# Audit Log → Export CEF downloads a .cef file for SIEM ingestion.

# Format:
CEF:0|Cheqpoint|ApprovalGate|1.0|<id>|<action>|<severity>|\
  rt=<timestamp> suser=<reviewer> duser=<agent> act=<status> msg=<reason>

# Severity mapping:
#   riskScore 0.0–0.3  → 3 (Low)
#   riskScore 0.3–0.7  → 6 (Medium)
#   riskScore 0.7–1.0  → 9 (High)

# Ingest into Splunk, Sentinel, Elastic, or any SIEM via file-based collector.

AI agent threat landscape

AI agents take real-world actions - sending emails, executing database writes, processing refunds, modifying infrastructure. Each of those actions is an attack surface. Cheqpoint adds a mandatory human gate before any high-stakes action executes.

Prompt injection

An attacker embeds malicious instructions in data an agent reads - a support ticket, a web page, a document - causing the agent to take an action it was never intended to take (exfiltrate data, send an email, escalate privileges). Cheqpoint intercepts the action before it executes, surfacing it for human review with the full payload visible. A reviewer sees the anomalous action and rejects it.

Runaway agents

Agents with tool access can take chains of actions autonomously. A bug, an unexpected edge case, or a misconfigured prompt can cause an agent to loop, escalate scope, or take destructive actions at scale before anyone notices. Cheqpoint's approval gate halts the action chain at any point you define. The kill switch suspends all pending and incoming requests from an agent immediately.

Scope creep

Agents instructed to do one thing will sometimes attempt actions outside their intended scope - not through malice, but through ambiguity in their system prompt or unexpected tool combinations. Cheqpoint rules let you define which actions always require approval and which can be auto-approved, creating an explicit security policy around each agent.

Unverified agent actions at scale

Without an approval gate, AI agents can process thousands of actions per hour with no human in the loop. A single bad decision - one incorrect refund rule, one misconfigured policy - propagates at machine speed. Cheqpoint gives you a circuit breaker: high-risk actions pause for review, low-risk actions auto-proceed after you've explicitly approved that pattern.

Data exfiltration via agent

An agent with access to a CRM, database, or file system can be manipulated into sending sensitive data to an external destination. Cheqpoint surfaces the action - including the full payload - before it executes, making exfiltration attempts visible to a human reviewer before they succeed.

How Cheqpoint works as a security control

Mandatory approval gate - actions are blocked by default until a human approves. The agent's execution thread blocks synchronously. There is no way for the agent to bypass the gate once checkpoint() is called.
Full payload visibility - reviewers see the complete action payload in Technical View, not just a summary. This makes it possible to spot manipulated or anomalous values.
Anomaly detection - the dashboard surfaces agents with request volume ≥ 2× their 30-day baseline, or any agent with a riskScore > 0.8 in the last 24 hours.
Kill switch - any agent can be suspended instantly from the dashboard or via API, halting all pending and incoming requests immediately.
Immutable audit log - every request, decision, and modification is permanently recorded with timestamps and reviewer identity. Exportable as CSV or CEF for SIEM ingestion.
Auto-approval rules as security policy - explicitly define what agents are allowed to do without review. Anything outside that policy requires human sign-off.
Risk scoring - pass a riskScore (0.0–1.0) with each request to surface the highest-risk actions first and trigger alerts in connected SIEMs.

SIEM integration

Cheqpoint forwards approval events to your SIEM in real time via Notification URLs (webhooks). CEF export is available for file-based ingestion. Supported platforms:

DatadogSplunkMicrosoft SentinelElastic SecurityCrowdStrike Falcon LogScale

See the Security & SIEM docs section for per-platform setup guides.

Cheqpoint infrastructure security

Data residency

All data is stored in the London, UK region (AWS eu-west-2). Request payloads are retained only for the duration of your workspace retention policy.

Encryption

All data encrypted in transit via TLS 1.3 and at rest via AES-256. Encryption keys managed with AWS KMS.

Connection Keys

Every AI Assistant authenticates via a cryptographically secure Connection Key (cq_live_...). Keys are stored hashed, scoped to specific agents, and can be rotated instantly if a leak is suspected.

Compliance

GDPR compliant. SOC 2 Type II certified infrastructure. DPAs available for Enterprise customers. Cheqpoint approval records satisfy EU AI Act Article 14 human oversight documentation requirements. See the EU AI Act compliance guide.