How to report a security issue to us.
We build a defensive platform, so we take reports about its weaknesses seriously. This page covers what is in scope, how to reach us, the response times we hold ourselves to and safe harbour for good-faith research. It also covers what happens when disclosure goes outside these rules.
What you can test.
The Safeguard platform
Authentication, authorisation, tenant isolation, scan pipelines, the Marketplace, the API surface, the IDE extension, the CLI, the MCP server, and the public web properties on safeguard.sh and its subdomains.
The model family inference paths
Griffin, Eagle, Lion, and their hosted inference endpoints. Vulnerabilities in trace emission, attestation, refusal behaviour, or guardrail evaluation are in scope.
Customer-facing surfaces in air-gapped deployments
The air-gapped appliance and its update channel. If you run Safeguard air-gapped, vulnerabilities in the appliance itself are in scope under a separate co-ordinated channel — write to security@safeguard.sh for the appropriate path.
The supply chain that produces our binaries
Compromise paths through our build, signing, or release pipeline. The SBOMs and attestations we publish for our own binaries are themselves in scope.
What is not covered.
Findings without a demonstrated security impact
Missing security headers without an exploit chain, theoretical issues without a working proof of concept, or best-practice deviations that do not reduce to a CIA-triad impact.
Volumetric or denial-of-service testing
Rate-limit testing, flood traffic, brute-force credential stuffing, or any test whose primary signal is volume. Coordinate with us before any load-style assessment.
Social engineering of staff or customers
Phishing, vishing, pretexting, or any test targeting people rather than systems. Out of scope without prior written authorisation from our security team.
Third-party services we depend on
Vulnerabilities in upstream cloud providers, open-source components, or third-party SaaS we use should be reported to the relevant maintainer or vendor first. We will help with co-ordinated disclosure where appropriate.
How to report
Email security@safeguard.sh with a clear description of the issue, the steps to reproduce, the impact and any supporting artefacts. Use English where possible. A PGP key is available on request. Ask in your first message and we will send the key and a fingerprint over a separate channel.
Please do not file security issues on the public forum or in GitHub issues. Use the inbox.
Our SLA to you
- Acknowledgement
- Less than 1 business day
- Triage
- Less than 5 business days
- Fix
- Severity-dependent
Critical issues are remediated as a fire drill. Lower-severity issues land on the next scheduled release train. We update you weekly until the fix ships.
Recognition, not payment
We do not pay for vulnerability reports. There is no bug bounty, cash reward or other payment, and nothing on this site promises one.
What we offer is credit. Once the fix ships, a valid report is listed in our Hall of Fame under the name and link you choose, or not at all if you would rather stay anonymous.
Decoys and sandboxes
Some of what looks like Safeguard is a decoy: an isolated sandbox full of synthetic data, watched by our own agents, who are happy to play along. No decoy is connected to production or holds customer data.
Reaching a decoy is not a vulnerability, and everything done inside one is logged. Not sure whether you hit one? Report it anyway and we will tell you. Decoy hits can end up in the Caught in the sandbox list.
Safe harbour
We will not pursue civil or criminal action against researchers who, in good faith, comply with this policy. That includes: testing only against in-scope surfaces, avoiding privacy violations and service degradation, not exfiltrating data beyond what is needed to demonstrate impact, keeping the issue confidential until we have shipped a fix, and giving us a reasonable opportunity to remediate before public disclosure.
If you are unsure whether a planned test crosses a line, ask first. We are reachable, and we would rather have a five-minute conversation than retroactively read about it.
Disclosure outside this policy.
Safe harbour covers good-faith research done under this policy. It does not cover the conduct below, and Safeguard reserves every remedy the law gives it.
Unlawful public disclosure
Publishing, selling or sharing with anyone other than Safeguard the details of a vulnerability, or of a claimed one, before we have shipped a fix or agreed a disclosure date with you in writing is outside this policy, and safe harbour does not apply to it. Where that disclosure, or the testing behind it, breaks the law, the person responsible may be held liable for the harm it causes to Safeguard, our customers or their users.
Defamation in any medium
False statements of fact about Safeguard's security are not research. That includes presenting a decoy or sandbox hit as a breach of our platform or of customer data. It applies to every form of media and communication: social media, blogs, videos, podcasts, forums, the press, messaging apps, talks and email. Anyone who makes such statements may be held legally liable under the defamation, trade libel or similar laws that apply where they are published.
The law of the country where it happens
Conduct covered by this notice is subject to the laws of each country where it takes place or has effect, including laws on computer misuse, data protection and defamation. Safeguard reserves the right to enforce its rights under those laws, to pursue the remedies they provide and to refer matters to the authorities in that country. This is in addition to our Terms and Conditions.
None of this limits a truthful statement or an honest opinion made in good faith, disclosure on a date we have agreed with you, or a report you make to a regulator or law enforcement. If you are not sure whether something you plan to publish is covered, email security@safeguard.sh first.
Credit where credit is due.
Researchers whose reports lead to a fix are credited in our Hall of Fame, unless they ask to stay anonymous, which we honour. The same page carries honourable mentions and the decoys people have mistaken for the real thing.
View the hall of fameRelated
The wider surface area for security and the commitments we hold ourselves to.
Have a finding that needs a live conversation?
For critical issues that warrant a real-time discussion, book a slot with our security team. We will set up a bridge within hours.
Self-healing security runs on Safeguard.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.