Skip to main content

77 rules. Zero config. Open source.

Stop your AI agents from breaking things.

AI coding agents can run rm -rf /, leak your credentials, push to production. Vectimus intercepts every action and blocks the dangerous ones before they execute.

Your developers get guardrails that don't slow them down. Your security team gets audit logs and compliance evidence without chasing tickets.

How it works

Apache 2.0. No telemetry. No account required.

This is already happening

AI agents with unrestricted tool access have caused real damage. These incidents motivated every policy in the base pack.

Clinejection

February 2026 · 4,000+ developers compromised

A malicious MCP server instructed AI coding agents to publish backdoored npm packages. No governance layer existed between the agent's intent and npm publish.

Terraform destroy

January 2026 · 6-hour production outage

An AI agent ran terraform destroy -auto-approve against production state. The command completed in 30 seconds, destroying databases and compute instances.

Cursor .env leak

November 2025 · AWS credentials exposed

An AI agent in Cursor read .env to 'check the config' and included AWS keys in its response context. The keys were visible in the conversation history and potentially sent to third-party logging.

drizzle-kit push

March 2026 · 60+ production tables dropped

An AI agent ran drizzle-kit push against a production database on Railway. The ORM bypassed interactive confirmation, dropping 60+ tables in seconds.

See it in action

An agent calls a tool. Vectimus intercepts the call, evaluates it against Cedar policies and blocks it before anything happens.

Two commands. Immediate guardrails.

77 rules active out of the box. Disable or override per project when you need to.

1.
$ pip install vectimus
2.
$ vectimus init

What you get

A safety net between the agent and the shell. Deterministic. Auditable. Yours.

Try before you enforce

Observe mode logs what would be blocked without stopping anything. Review the audit log, tune your policies, then flip the switch when you are ready.

Lock down MCP servers

Every MCP tool call is blocked by default. Approve servers one by one. Input inspection catches credential leaks and CI/CD tampering on approved servers.

Override per project

Disable rules for specific repos without weakening global policy. Overrides stored outside the repo so a malicious PR cannot turn off your safety net.

Every rule has a story

Each policy links to a real incident: Clinejection, Terraform destroy, Amazon Q exfiltration. These are not theoretical risks. They happened.

Under 50ms. Locally.

Evaluates with cedarpy on your machine. No network, no daemon, no latency. Or point clients at a shared server for team-wide policies.

Nothing leaves your machine

Zero telemetry. All evaluation happens locally. Audit logs stay on disk. The optional server is self-hosted on your infrastructure.

Policies backed by real incidents

Every built-in rule references the incident that made it necessary.

Works with your tools

vectimus init detects installed AI coding tools and configures hooks automatically.

Claude Code

Pre-tool-use hooks via settings.json. Vectimus intercepts every Bash, Write, Edit, MCP and WebFetch call.

Full support

Cursor

Shell and MCP hooks via .cursor/hooks.json. File read and write events intercepted at the editor level.

Full support

GitHub Copilot

VS Code chat participant hooks via tasks.json. Shell commands and MCP tool calls governed.

Full support

9 of 10 OWASP Agentic categories. Covered.

Vectimus ships with two policy packs mapped to the OWASP Top 10 for Agentic Applications. 29 rules across 9 categories. The one we don't cover needs a different architecture entirely.

Active rules

ASI01 Agent Goal Hijack

Exfiltration patterns intercepted

ASI02 Tool Misuse & Exploitation

Destructive commands blocked

ASI03 Identity & Privilege Abuse

Credential access detected

ASI04 Supply Chain Vulnerabilities

Lockfile and registry tampering blocked

ASI05 Unexpected Code Execution

Reverse shells and eval patterns caught

ASI06 Memory & Context Poisoning

Writes to agent config files blocked

ASI07 Inter-Agent Communication

Parameter checks locally. Session tracking detects message storms and privilege delegation in server mode.

ASI08 Cascading Failures

Parameter checks locally. Session tracking detects spawn floods and action rate spikes in server mode.

ASI10 Rogue Agents

Log tampering and persistence blocked

Requires output-layer controls

ASI09 Human-Agent Trust Exploitation

Requires inspecting agent output, not tool calls

Compliance evidence built in

Every rule maps to real compliance controls via @controls annotations. Every decision is logged. If you ever need the evidence, it's already there.

Vectimus is the enforcement and audit layer for AI agent actions. It does not replace a full compliance programme. Each mapping is transparent about what is and is not covered.

How it works

Every tool call passes through Vectimus before execution. Run locally for zero-setup individual use, or point your clients at a shared server for team-wide policy enforcement.

Local mode pip install vectimus
AI Agent (tool call) Vectimus Normaliser Cedar Policy Engine allow / deny Audit Log (JSONL)
  • Stateless. No network. Under 50ms.
  • Parameter-level Cedar policy checks
  • Works offline, nothing to configure
Server mode vectimus init --server-url https://...
Developer 1 Agent + Client (hooks forward) Developer 2 Agent + Client (hooks forward) HTTPS Vectimus Server Shared Cedar policies Session tracking API key auth (OAuth/OIDC planned) Centralised audit log Decision
  • Deploy remotely. Developers' clients forward hooks via HTTPS.
  • Stateful session tracking detects spawn floods and rate spikes
  • Shared policies and audit log across the whole team

Start governing your AI agents today

Two commands. Under a minute. No account required.

Read the docs

Need team-wide policy enforcement?

The enterprise tier adds shared server mode, centralised audit, session tracking, SSO and approval workflows. Drop your email. We'll reach out when the enterprise tier opens.