Vigil is a security platform. We hold ourselves to the standard we ask of the tools we integrate with: report vulnerabilities privately, get a human response quickly, and coordinate disclosure so operators can patch before details are public.
The machine-readable version of this policy is published per RFC 9116 at https://vigilsoc.org/.well-known/security.txt.
Vigil is pre-1.0. Security fixes land on main and ship in the next release —
we do not backport to earlier minors.
| Version | Supported | Notes |
|---|---|---|
Latest released 0.x minor (latest release) |
✅ | Fixes ship here. Upgrade before reporting a bug you can only reproduce on an older tag. |
Older 0.x minors |
❌ | No backports. What may change between releases is in VERSIONING.md. |
main (unreleased) |
✅ | Reports welcome; note the commit SHA. |
Container images (ghcr.io/vigil-soc/vigil-backend, vigil-daemon, vigil-agent,
vigil-mcp-runner) and the Helm chart at infra/helm/vigil/ follow the same window.
Chart version and appVersion move in lockstep, so the chart version
identifies the app release it deploys.
Release images are signed keyless by .github/workflows/release.yml (GitHub
OIDC → Fulcio, Rekor). Confirm an image came from that workflow — the identity
is pinned to this file and issuer, not any Actions run in the repo. Image tags
drop the leading v; the certificate identity uses the git tag:
cosign verify \
--certificate-identity https://github.com/Vigil-SOC/vigil/.github/workflows/release.yml@refs/tags/v<version> \
--certificate-oidc-issuer https://token.actions.githubusercontent.com \
ghcr.io/vigil-soc/vigil-backend:<version>The same invocation works for vigil-daemon, vigil-agent and vigil-mcp-runner.
Do not open a public GitHub issue, pull request, or discussion for a security vulnerability. Do not disclose it on social media, a blog, or a conference talk before the coordination window below has closed.
Use whichever channel fits you — both reach the same maintainers.
https://github.com/Vigil-SOC/vigil/security/advisories/new
This is the fastest path. It gives us a private fork to develop and review the fix in, keeps the whole thread with the eventual advisory, and lets us request a CVE and credit you in one step. You need a GitHub account; nothing else.
security@deeptempo.ai — the point of contact for anything that should not or cannot go through the GitHub advisory process: you have no GitHub account, the report concerns Vigil's hosted or website infrastructure rather than this repository, you need to reach a human about an in-flight report, or GitHub itself is the wrong venue for what you found.
If you need to encrypt, say so in a first message with no technical detail and we will reply with a key. Please do not send unsolicited encrypted mail to an address you have not confirmed we can read.
A report we can reproduce is a report we can fix in days instead of weeks:
- Impact — what an attacker gains (auth bypass, RCE, secret disclosure, privilege escalation, SSRF via an integration, prompt injection into an agent with tool access, etc.).
- Affected component — path and, ideally, line:
services/api/,core/llm/harness/, an MCP server undercore/integrations/<vendor>/,clients/web/, aninfra/database/init/migration, the Helm chart. - Version — release tag, commit SHA, or image digest.
- Reproduction — minimal steps, a request/response pair, a curl invocation,
a proof-of-concept, or a failing test. Include your
DEV_MODEsetting: withDEV_MODE=trueall authentication is bypassed by design, and findings that depend on it are not vulnerabilities (see Out of scope). - Deployment shape — docker-compose, Helm, desktop app, or bare
uvicorn. - Attribution — how you want to be credited, or that you prefer to stay anonymous.
Please keep your testing to your own deployment. Do not test against
vigilsoc.org, another organization's Vigil instance, or any third-party
service an integration talks to.
| Stage | Target |
|---|---|
| Acknowledgment that a human has your report | 3 business days |
| Triage: confirmed or rejected, with severity | 7 business days |
| Fix or documented mitigation for critical/high | 30 days from triage |
| Fix or documented mitigation for medium/low | Next scheduled release |
| Public advisory | Coordinated with you — see below |
These are targets, not contractual guarantees; Vigil is maintained by a small team. We will tell you when something is going to slip rather than let the thread go quiet.
Throughout, we will: confirm receipt, tell you whether we accept the report and at what severity, keep you updated as the fix moves, and let you review the advisory text and your credit line before it is published.
We ask you to: give us a reasonable window before disclosing, not access or modify data that isn't yours, and tell us if you believe the issue is being exploited in the wild — that changes our timeline immediately.
We use CVSS v3.1 to set severity. Our rating is the one that drives our timeline, and we will explain it if it differs from yours.
Our default window is 90 days from triage to public disclosure, or the day the fix ships — whichever comes first. We publish sooner when a fix is ready sooner, and we would rather ship early than sit on a patch.
We will ask to extend past 90 days only when a fix is genuinely hard (a schema migration, a breaking interface change, an upstream dependency we don't control), and we will say why. If we go quiet for 90 days with no fix and no explanation, disclose — that is our failure, not yours.
Fixes are published as GitHub Security Advisories on the upstream repository. For anything of medium severity or above we request a CVE through GitHub, note the advisory in CHANGELOG.md, and name the affected and fixed versions so operators can tell whether they need to act. Watch that page, or the repository's releases, if you run Vigil in production.
If you make a good-faith effort to follow this policy, we will not pursue or support legal action against you, and we consider your research authorized under the Computer Fraud and Abuse Act and equivalent laws. If a third party brings action against you for research that followed this policy, we will say publicly that it was authorized.
Good faith means: you test only against deployments you own or have written permission to test, you stay within the scope below, you avoid privacy violations, data destruction, and any degradation of service, you access only the minimum data needed to demonstrate the issue, you delete any retained data once the report is closed, and you report promptly and give us the coordination window.
This is not a paid bug bounty. Vigil is an Apache-2.0 open-source project with no bounty budget; we pay in credit, in acknowledgment, and in fixing what you found.
Anything in this repository that a deployment actually runs:
- Authentication and authorization — JWT handling in
core/, the middleware underservices/api/middleware/, RBAC, session and cookie handling, token revocation, CSRF. - API surface — every router discovered by
services/api/discovery.py: injection, SSRF, path traversal, IDOR, missing authorization, mass assignment, unsafe deserialization. - The agent and LLM layer —
core/llm/,core/agents/. Prompt injection that escalates into real capability (unapproved tool execution, exfiltration of secrets or another tenant's case data, bypass of the response-approval gate) is in scope. Getting an agent to produce impolite or wrong text is not. - MCP integrations —
core/integrations/<vendor>/,tools/mcp/,mcp-config.json, andcore/integrations/mcp/service.py: credential leakage into logs or agent context, command injection into a spawned server, TLS verification that can be silently disabled, and bypass of the approval workflow incore/response/approval_service.pyfor actions that change the world (host isolation, firewall rules). - Secret handling — the encrypted store at
~/.vigil/secrets.enc,core.secrets, Helm secret plumbing, and anything that writes a credential to a log, an error response, or the database in cleartext. - Data layer — SQL injection, the
infra/database/init/migrations, and tenant or case isolation failures. - Web client —
clients/web/: XSS, CSRF, token storage, and anything that leaks one user's data to another. - Deployment artifacts — the Dockerfiles and compose file under
infra/docker/, the Helm chart, andscripts/: privilege escalation inside a container, a chart default that exposes an unauthenticated service, a secret baked into an image layer. - Supply chain — a dependency with a known exploitable CVE reachable from Vigil's own code paths, or a compromised build or release step.
Not vulnerabilities. Reports on these will be closed with a pointer back here:
DEV_MODE=truebypassing authentication. That is its documented purpose. It defaults tofalseincore/config.pyand inenv.example; a developer opts into it for a local instance, and the backend announces the bypass on every startup. The backend refuses to start with it whenENVIRONMENT=production, which the Helm chart sets by default. See DEV_MODE.- Default credentials in development material — the default PostgreSQL
password in
infra/docker/docker-compose.yml, documented as must-change-before-production in production security. No default admin login ships; the first account is created through/api/auth/bootstrap, which requires the setup code the API prints at startup while the user table is empty. - Placeholder values in
env.example. They are a template, not a configuration. - Findings that require a misconfiguration we already document as unsafe —
for example exposing port 6987 to the internet with
DEV_MODE=true, or running with the report-only CSRF switch turned on in production. If you find a case production security does not cover, that gap is worth reporting. - Vulnerabilities in an upstream product Vigil integrates with (Splunk, CrowdStrike, VirusTotal, and the rest). Report those to that vendor. If Vigil's client for one is what mishandles the credential or the response, that is in scope.
- Automated scanner output with no demonstrated impact, missing hardening headers with no exploit path, best-practice advice, outdated dependencies with no reachable vulnerability, and self-XSS or clickjacking on pages with no sensitive action.
- Denial of service through resource exhaustion or volumetric traffic, and
social engineering of maintainers or users. LLM cost-amplification that
defeats the guardrails in
services/daemon/orchestrator.pyis in scope — it costs operators real money.
Most incidents we would expect to see are configuration, not code. Before you run Vigil anywhere real:
- Production security — the auditable checklist of every security-relevant switch and its production value. Start here.
- DEV_MODE — what the auth bypass does and why it must never be enabled in production.
- State and secrets — where secrets live, and why provider
keys and integration credentials belong in the UI and the encrypted store
rather than in
.env. - Helm secrets — secret management for Kubernetes deployments.
- Deployment guide — network exposure, TLS termination, and reverse-proxy placement.
Non-negotiables: set DEV_MODE=false, generate a real JWT_SECRET_KEY, change
the default database password, terminate TLS in front of the API, and never
commit a credential.
Contributions are held to the same bar, enforced in CI (ci-cd.yml, plus the
nightly audit in nightly.yml):
banditfor Python andnpm auditfor the web client.- Trivy scans of released container images.
- Dependabot for dependency updates (
.github/dependabot.yml). lint-importsenforcing the layering in.importlinter.
Configuration and credentials are read through core.config or
core.secrets, never a new os.getenv call, which keeps credentials out of
scattered env reads.
If you are fixing a reported vulnerability, coordinate with the maintainers before opening a public pull request — a PR that describes the bug discloses it. We will work with you in the private advisory fork instead.
This document is the human-readable half of the policy published at
https://vigilsoc.org/.well-known/security.txt. The two must agree; when they
disagree, this file is authoritative and security.txt should be corrected.
security.txt field |
Corresponds to |
|---|---|
Contact (advisory URL) |
GitHub private vulnerability reporting |
Contact (mailto:) |
security@deeptempo.ai — see Email |
Policy |
This file |
Acknowledgments |
Acknowledgments & Hall of Fame |
Preferred-Languages |
en |
Canonical |
https://vigilsoc.org/.well-known/security.txt |
Expires |
Reviewed and re-signed before it lapses |
Maintainers: any change to a contact address, the disclosure window, or the
section headings this table anchors to must be mirrored in security.txt in
the same change. The Acknowledgments field points at the
#acknowledgments--hall-of-fame anchor below, so renaming that heading breaks
the published link. Refresh Expires on every review — an expired
security.txt is treated as an unmaintained policy.
Researchers who have reported valid vulnerabilities to Vigil and coordinated disclosure with us. We add you here when the advisory publishes, with the name or handle you asked for — or not at all, if you would rather stay anonymous.
No entries yet. This section is intentionally empty, not neglected — Vigil is young. Be the first.
Thank you for helping keep Vigil, and the teams who run it, safe.