CadreGPT AI Safety Standards & Policy

Acceptable Use Policy & AI Safety Rules

Mandatory rules of engagement governing autonomous agent execution, API tool usage, and commercial foundation model consumption across CadreGPT.

Effective: October 2026•Applies to: All Workspaces, Agent Topologies & API Tokens•Strict Enforcement Active

Protecting Enterprise Foundation Model API Quotas & Master Keys

CadreGPT maintains high-throughput commercial enterprise API tiers with Google Cloud, Anthropic, and OpenAI. To safeguard the shared platform reputation, comply with upstream provider acceptable use policies, and guarantee 99.9% availability for all enterprise tenants, we enforce automated real-time heuristic safety filters and zero-tolerance policy measures against abusive agent workflows.

Strictly Prohibited Autonomous Agent Activities

Users, developers, and organizations may not direct, program, or prompt CadreGPT agents or custom workflows to engage in any of the following activities:

Automated Spam & Phishing

Sending unsolicited bulk email, cold outreach spam, or deceptive marketing campaigns.
Generating phishing lures, credential harvesting forms, or social engineering pretexts.
Conducting aggressive, unauthorized web scraping that disregards target robots.txt directives or terms of service.

Exploit Generation & Jailbreaking

Crafting prompt injection attacks, "jailbreak" adversarial vectors, or filter bypasses against foundation models.
Developing functional malware, ransomware, botnet exploits, or automated cyber attack payloads.
Scanning third-party network ports or testing targets without verifiable, prior written authorization.

Infinite Loop & Compute Abuse

Executing deliberate runaway loops designed to flood upstream model gateways or consume excessive compute quotas.
Attempting to bypass platform rate limiters, token metering bounds, or concurrency gates.
Using sandboxed worker runtimes for cryptocurrency mining, distributed password cracking, or denial-of-service (DDoS).

Unstructured Regulated Data

Ingesting Protected Health Information (HIPAA PHI) into unstructured prompts without an executed Enterprise Business Associate Agreement.
Transmitting raw Primary Account Numbers (PCI-DSS PAN/CVV) into agent text streams or tool arguments.
Uploading unredacted government IDs, social security numbers, or biometric identification databases.

Harmful & Unlawful Content

Generating child sexual abuse material (CSAM), non-consensual synthetic intimate imagery (deepfakes), or suicide promotion.
Promoting violent extremism, terrorism recruitment, or actionable weapons manufacturing schematics.
Fabricating defamatory, slanderous, or intentionally deceptive impersonations of living persons or regulatory authorities.

High-Stakes Unsupervised Actions

Deploying fully unsupervised autonomous agents for legal judgments, credit eligibility scoring, or hiring termination decisions.
Bypassing mandatory Human-in-the-Loop gates on high-privilege external database mutations or live payment transactions.
Operating critical infrastructure (medical equipment, traffic systems) with non-deterministic agent control.

ACTIVE PLATFORM DEFENSE

Multi-Layered Gateway Tripwires & Anomaly Detection

CadreGPT’s Central Model Gateway evaluates all agent prompt expansions and tool parameters through automated heuristic tripwires before dispatching requests to upstream foundation model APIs:

Prompt Injection Tripwire

Regex and semantic classifiers flag prompt hijacking or role-reversal attempts before model inference.

Velocity & Loop Circuit Breakers

Dynamic cycle detectors identify repetitive tool invocation sequences and pause recursive loops.

Credential Vault Isolation

Raw OAuth tokens and API secrets are never injected into prompt contexts, preventing prompt exfiltration.

Tripwire trigger events log immediate security alerts to the CadreGPT Trust & Safety Engineering Operations team.

Enforcement Protocols, Heartbeat Suspension & Account Termination

CadreGPT reserves the right, at its sole discretion, to investigate suspected violations of this Acceptable Use Policy and take any of the following enforcement measures:

1. Immediate Agent Heartbeat Suspension:
The BullMQ task worker immediately cancels the running job, aborts pending external tool calls, and marks the task as TERMINATED_SAFETY_POLICY.

2. API Key Invalidation & Workspace Lock:
Workspaces engaging in persistent prompt injection, phishing generation, or credential harvesting are locked. All integration secrets in the envelope vault are disabled.

3. Forfeiture of Compute Credits:
In accordance with our Terms of Service, accounts terminated due to intentional, malicious violations of the AUP forfeit all remaining Agent Credits and subscription fees without refund.

4. Law Enforcement & Upstream Reporting:
Illegal activities (including child exploitation, unauthorized cyber attacks, and terrorist financing) will be reported directly to law enforcement authorities and upstream foundation model providers (Google, Anthropic, OpenAI).

Believe a task was flagged in error? Submit an appeal to: [email protected]

Contact Trust & Safety Team