Operations
Responsible Disclosure
How to report a security vulnerability in CogniWork, what we promise in return, and what is in and out of scope.
If you have found a security problem in CogniWork, we want to hear about it, and we will not come after you for telling us.
Report to: [email protected]
Safe harbour
If you make a good-faith effort to comply with this policy while researching a vulnerability, we will:
- consider your research authorised under the Computer Fraud and Abuse Act and equivalent laws;
- consider it exempt from the restrictions in our Acceptable Use Policy and Terms of Service that would otherwise prohibit it;
- not pursue or support legal action against you;
- work with you if a third party takes action against you for research conducted under this policy.
If you are unsure whether something is in scope, ask us first at [email protected]. We would much rather answer the question than have you guess.
What we ask
- Use your own accounts and data. Create a free account and test against your own workspace.
- Stop at proof. Once you have demonstrated a vulnerability, stop. Do not enumerate further, exfiltrate data, or pivot.
- Don’t touch other people’s data. If you encounter data belonging to someone else, stop immediately, do not save it, and tell us what you saw so we can assess exposure.
- Don’t degrade the service. No denial-of-service testing, no load testing, no automated scanning that generates significant traffic, no spam or social engineering of our staff or users.
- Don’t disclose publicly until we have had a reasonable chance to fix it — see the timeline below.
Timeline
| When | What we do |
|---|---|
| Within 2 business days | Acknowledge your report |
| Within 7 business days | Confirm the issue, or explain why we think it isn’t one, and give a severity assessment |
| Within 30 days | Fix critical and high-severity issues, or give you a dated plan |
| At fix | Tell you it’s fixed and, if you want, credit you |
| 90 days from report | Coordinated public disclosure, or sooner by agreement |
If we go quiet, chase us. If we still go quiet, you are free to disclose after 90 days.
In scope
- The CogniWork web application and its API
- The workspace MCP endpoint and its authorisation flows
- Authentication, session handling and workspace authorisation
- Isolation between workspaces — data, files, compute and integration credentials
- Agent-specific issues: privilege escalation through a worker, scope escalation on an MCP grant, credential leakage into transcripts or logs, prompt injection with a concrete security consequence such as unauthorised data access or a cross-workspace effect
Out of scope
- Findings from automated scanners without a demonstrated exploit
- Missing security headers, cookie flags or TLS configuration with no demonstrated impact
- Rate limiting on non-authentication endpoints
- Self-XSS, clickjacking on pages with no state-changing action, or issues requiring a compromised device
- Social engineering, phishing, or physical attacks
- Denial of service, resource exhaustion, or volumetric attacks
- Vulnerabilities in third-party services we merely connect to — report those to their owners
- Model behaviour alone. An LLM producing an undesirable, offensive or inaccurate output is not a security vulnerability. A prompt injection that causes unauthorised access to data, or crosses a workspace boundary, is — please report that.
What to include
- What the issue is, in a sentence.
- Steps to reproduce, precisely enough that we can follow them.
- What an attacker gets — the actual impact.
- Affected URLs, accounts or workspace identifiers.
- Anything else you think matters.
Screenshots and a short video help. Please encrypt anything sensitive; ask us for a key.
Rewards
We do not currently run a paid bug bounty. We do offer public credit — with your name or handle, or anonymously, your choice — and, for findings that materially improve the product’s security, something better than a thank-you. Ask.