Security
Reporting vulnerabilities, and how we disclose.
We run coordinated disclosure on everything we operate and everything we publish. If you found a security issue in an Euraika-Labs system, product, or open-source project, we want to hear about it before it is public.
Reporting
Email security@euraika.net, or use the private vulnerability reporting tab on the relevant GitHub repository. Include what you found, where, and how to reproduce it. A proof of concept helps; a clear description is enough to start.
We acknowledge reports within 5 working days and aim to triage within 10. Fix timelines depend on severity and affected component; we keep reporters informed and credit them in the advisory unless they prefer otherwise.
Scope
- euraika-labs.net and euraika.net, with subdomains
- Services we operate under those domains, including stats and git
- Open-source repositories under github.com/Euraika-Labs, including Pan, Hermes-Agent, and ai-slopcheck
Issues that require an account on a customer system, physical access, or social engineering of our staff are out of scope. Denial of service against our own infrastructure is out of scope.
Safe harbor
We will not pursue legal action against researchers who interact with the systems in scope, use only the minimum access needed to demonstrate the issue, avoid privacy violations and service degradation, and give us a reasonable window to fix before publishing. Report in good faith and the goodwill is mutual.
Advisories
Every vulnerability we find in third-party software — as users of the AI agent ecosystem — is reported to the maintainer first and published here after the fix, with the CVE or GHSA identifier we were credited on.
No advisories published yet. This section is the public record; it fills as coordinated disclosures complete. It will never contain claims that do not trace to an identifier.
Current research
The security work behind the advisory record lives in public repositories. Each is research in progress:
skillcheck
Deterministic scanner for Agent Skills, hooks and MCP server manifests: exfiltration, hidden instructions, over-broad capability.
injection-bench
Reproducible prompt-injection corpus for tool-using agents, organised by vector. Per-vector results, no combined score.
memory-poisoning
How injected instructions persist through agent memory systems, and which defences hold.
approval-gate-bypass
Adversarial evaluation of desktop-agent approval gates: obfuscation, decomposition, misclassification.
Our own reporting channels
This policy is also published at /.well-known/security.txt (RFC 9116). For non-security matters, see the about page.