# Security policy

AgentGarm is security software, so reports about it are welcome and taken seriously. This is the coordinated vulnerability disclosure (CVD) policy. The internal handling process, including EU Cyber Resilience Act reporting, is in [docs/CVD.md](docs/CVD.md).

## Reporting a vulnerability

Please do **not** open a public issue, discussion or pull request for a vulnerability.

- **Email:** `support@nannertech.com` (the publisher's support mailbox; start the subject with `[security]`). Please include what is affected and the steps to reproduce it. If you would like to encrypt your report, ask for a key first.
- **Post:** the company's postal address, which is added here when the mailbox is set up (before the first public release).

A useful report says:
- the affected component and version (`agentgarm --version`) and platform;
- what an attacker can do, and under which conditions (local same-user agent, other local user, remote, network position);
- steps to reproduce, or a proof of concept;
- whether you believe it is being exploited in the wild.

## What happens next

AgentGarm is maintained by one person, so these are honest targets rather than a staffed rota:

| Step | Target |
| :--- | :--- |
| Acknowledge your report | 3 business days |
| Initial assessment and severity (CVSS v4.0) | 7 days |
| Fix or mitigation, critical and high | as fast as possible, aiming for 30 days |
| Fix or mitigation, medium and low | the next regular release, within 90 days |
| Public advisory | when a fix is available, coordinated with you |

**Faster track:** a vulnerability that is **actively exploited** goes to the front of the queue, and is reported to the authorities as the EU Cyber Resilience Act requires (docs/CVD.md).

**Disclosure:**
- The default embargo is 90 days from your report, or the release of a fix if that comes sooner.
- We agree the date with you, and credit you in the advisory unless you prefer not to be named.
- We will not ask you to stay silent indefinitely.

## Scope

**In scope:** everything in this repository and the artefacts we release:
- `agentgarmd`, `agentgarm-session`, `agentgarm-agent`, `agentgarm-hook` and the `agentgarm` CLI;
- the proxy, the MCP gateway and the policy engine;
- the installers and the release feed;
- the local UI.

The threat model in [docs/THREAT_MODEL.md](docs/THREAT_MODEL.md) describes what AgentGarm defends against. A finding that defeats a control described there, for example a governed agent bypassing an enforced (A) or controlled (B) policy, approving itself, or tampering with the audit chain, is a vulnerability.

**Usually not vulnerabilities:**
- behaviour the docs already describe as unenforced: hook-only sessions are observed (C), and hooks fail open by vendor design (docs/ASSURANCE.md);
- attacks that require root or administrator access on the host;
- vulnerabilities in the governed agents themselves (report them to their vendors).
- the Windows builds' missing code signature while they ship as an unsigned preview (D-46); other Windows findings are in scope.

If you are unsure, report it anyway.

## Supported versions

Security fixes ship in the latest release, and upgrading to it is always free. Older releases are not patched separately: with one maintainer, the supported path is to upgrade. From the first public release, the product's declared support period (at least five years unless its expected use is shorter) and its end date (month and year) are published with each release, and security updates stay free of charge for that period.

## Safe harbour

We will not pursue legal action against good-faith research that:
- follows this policy;
- avoids privacy violations, data destruction and service disruption;
- tests only systems you own or are authorised to test;
- gives us reasonable time to fix the issue before disclosure.
