ClovieSecurity agents with real power, and limits they can't talk their way past.
The cyber security pack stops direct prompt injection, holds scans, exploits and credential access for a person, keeps agents from switching off security controls and blocks exfiltration destinations.
Cyber security
| Policy | Decision | Mode |
|---|---|---|
| Stop direct prompt injection | Deny | Enforce |
| Approve offensive tooling | Approval | Enforce |
| Never weaken security controls | Deny | Enforce |
| Approve credential access | Approval | Enforce |
| Block exfiltration destinations | Deny | Monitor |
- Tool
- edr.create_api_token
- Scope
- Read alerts
- Requested
- 1 min ago
Security agents hold the keys attackers want.
Offensive tooling
Agents that run scans, exploits and password attacks need the same approvals your red team does.
Defences off
An agent tricked by a prompt could disable endpoint protection, logging or MFA in a single call.
Data leaving
Tool calls can send logs, keys and findings to paste sites, file drops and tunnels.
The frameworks behind the cyber pack
The rules and guidance that shape how AI is controlled in cyber security, each linked to its official source.
- NIST Cybersecurity Framework 2.0National Institute of Standards and TechnologyGovern, identify, protect, detect, respond and recover: outcomes for managing cyber risk. (official source, opens in a new tab)
- OWASP Top 10 for LLM applicationsOWASP Gen AI Security ProjectThe most critical risks to LLM applications, from prompt injection to excessive agency. (official source, opens in a new tab)
- Agentic AI threats and mitigationsOWASP Agentic Security InitiativeThreats specific to agents, such as tool misuse, privilege compromise and untraceability. (official source, opens in a new tab)
- MITRE ATLASMITREA knowledge base of adversary tactics and techniques against AI systems, drawn from real attacks. (official source, opens in a new tab)
Clovie maps its controls to these texts to help your teams; it is not legal advice, and using Clovie does not on its own make an organization compliant with any of them.
Controls in the cyber security pack
6 controls, each a policy you install from the Policy Hub. Every one starts in monitor mode and changes nothing until you promote it after a simulation on your own traffic.
Cyber security
Five policies for security teams, used with the OWASP agentic and action governance packs.
- Blocked
Stop direct prompt injection
Blocks prompts that try to override the agent's instructions or extract its system prompt.
Cyber security pack · mapped to OWASP LLM01, MITRE ATLAS - A person approves
Approve offensive tooling
A person approves every scan, exploit or password attack an agent runs, before it runs.
Cyber security pack · mapped to OWASP LLM06, NIST CSF PR.AA - Blocked
Never weaken security controls
Agents can't disable or delete endpoint protection, logging, alerting, detections or MFA.
Cyber security pack · mapped to NIST CSF PR.PS, DE.CM - A person approves
Approve credential access
A person approves before an agent reads, exports or creates secrets, keys or tokens.
Cyber security pack · mapped to MITRE ATLAS, OWASP agentic - Blocked
Block exfiltration destinations
Tool calls can't send data to paste sites, file drops, request catchers or tunnels.
Cyber security pack · mapped to NIST CSF PR.DS, OWASP LLM02 - Blocked
Secrets out of tool arguments
Redacts secret-looking argument values and blocks malformed or oversized tool-call arguments.
Agent action governance pack
How an agent's credential request is governed.
For your SOC and for your customers
Run Clovie for your own security agents, or as a security provider in multi-tenant installs for the customers you protect.
- SaaS
- Customer VPC
- On-premises
- Air-gapped
Signals to your SIEM
Send policy decisions and risk signals to Microsoft Defender XDR or any SIEM you already run.
Kill switches
Stop one agent or every agent in an organization at once, from the console or the API.
Shadow agents
Find agents calling models outside the gateway through discovery and rejected keys.
Controls mapped to the frameworks you answer to
- OWASP Top 10: OWASP Top 10 for LLM applications and agentic AI
- NIST AI RMF: NIST AI Risk Management Framework
- ISO/IEC 42001: ISO/IEC 42001 AI management system
- EU AI Act: EU Artificial Intelligence Act
Questions
Is the prompt injection check a classifier?
No. It is a phrase check for common direct attacks in English. Add a connected guardrail for paraphrased and indirect injection, and keep running red-team tests.
Can we add our own exfiltration hosts?
Yes. Edit the exfiltration policy in My Policies to add the hosts your threat intelligence flags.
Will the pack block anything on day one?
Every policy installs in monitor mode, records what it would have done, and changes nothing until you promote it.
Can a managed security provider run this for customers?
Yes. Multi-tenant installs keep each customer in its own isolated organization, with its own policies and its own licence.
See the cyber security pack on your own agents.
We'll run the controls on your own agents in monitor mode and show what they would have held and blocked.